RFD 1166 details: the table, the runoff, and what each row rests on
The STAR rank
STAR applied repeatedly: take the two highest sums, let the six dimensions vote between them, seat the winner, remove it, repeat. runoff reads wins-losses-ties against the runner-up it beat. loops is which of RFD 1143 to RFD 1146 the model appears in.
# model fit shp ref clr val adp ask wnt sum runoff role
1 rf-detr keypoint 100 100 100 80 100 95 100 85 760 6-0-2 fit
2 Qwen3-ASR-1.7B 95 85 45 80 80 75 95 85 640 4-3-1 ears
3 cyclegan_style_transfer 95 95 60 50 70 95 80 60 605 4-4-0 style, both
4 Qwen3-TTS CustomVoice 95 25 15 80 70 70 90 80 525 4-3-1 voice, host
5 LaMa, AnimeMangaInpaint 95 5 5 85 65 80 90 70 495 4-3-1 inpaint, host
6 OmniGen2 50 50 15 50 95 75 100 55 490 4-3-1 2D edit only
7 EditScore, Qwen3-VL-8B 40 45 90 55 60 70 100 70 530 6-2-0 all three
8 Kimodo 95 40 10 55 40 85 30 80 435 5-3-0 -
9 MoGe 85 90 55 50 30 60 40 50 460 4-4-0 bootstrap only
10 SkinTokens 95 15 10 45 35 80 30 85 395 4-3-1 -
11 MuJoCo MJX 95 30 5 15 20 90 85 80 420 3-3-2 presence
12 TRELLIS.2 / Pixal3D 60 35 5 25 75 30 35 90 355 4-3-1 3D backbone
13 Mitsuba 3 shading 95 75 5 10 55 15 95 45 395 4-4-0 3D views
14 VoxHammer 30 50 55 15 60 10 30 80 330 last 3D ingest+editThe sum is not the order. MoGe outscores EditScore by 20 and sits below it, having lost the runoff that decided the seat.
TRELLIS.2 and Pixal3D are one row. They were never independent candidates: Pixal3D’s README says it is “heavily built upon Trellis.2 … the foundation of our codebase and model architecture”, and its install step begins by following TRELLIS.2’s. Scoring them separately counted one backbone twice and let the derived model rank below the thing it is made of. The row takes the staged view RFD 1154 arrived at – the 1.3 B sparse structure stage rather than the 12.02 B whole – which is why fit is 60 rather than Pixal3D’s old 30.
Gemma 4 and qwen35-defiant are both struck from the ranking. Neither was a close call. Both are GGUF, so the format ends them before memory or operators are reached, and no exporter can close that gap because GGUF carries no graph at all. RFD 1155 abandoned Gemma 4 on the same ground.
Ranking a candidate with no path flatters the ranking rather than informing it, and keeping one as a worked example only made the table look like it had considered language models on their merits. It had not: it had rediscovered the blocklist. BLOCKLIST.md is where that argument lives.
unified-modal-embedder and residual-fsq-recommender are struck for a different reason: they are not what this product is. One embeds content for search and the other recommends a next item. Both are infrastructure for finding things, and the two wants this ranking serves are making and being present. Neither is in a chain, neither has run here, and the embedder held the lowest wanted of any surviving row at 20.
Nothing is wrong with either. A ranking of accelerator candidates is not a list of the workspace’s models, and the checkouts stay in the manifest where they are placed.
MoGe was struck and is restored, scoped to bootstrap. It is not an independent candidate: it sits in Pixal3D’s core requirements.txt and app.py calls it from get_camera_params_wild_moge, which takes a photograph somebody else shot and estimates the camera that shot it, producing exactly the camera_params dict run() demands.
The strike said the inverted chain deletes that input, and it does – after the first turn. It does not answer where the first asset comes from. ANNY originating a pose has no wild image and needs no camera recovered. A photograph does, and image-to-3D is a catalog entry with two models behind it, which is not what gets built if nobody brings images.
So the strike was wrong twice over. The case survives at the bootstrap, and more to the point, a dependency that might be needed is still code to write and QA on the device. This ranking tracks work the accelerator programme takes on, and work that turns out to be unnecessary is discovered by doing it. Its value falls from 40 to 30 because it runs once at ingestion rather than every round, and that is the whole of what the inversion changes.
wanted, and what the product is for
The other seven dimensions all ask whether a model is tractable. None asks whether the thing it does is wanted, and a model can be easy to accelerate and serve nothing anybody needs.
What the product is for answers it, and the manifest says: a social world with avatars people wear. v-sekai and v-sekai-fabric carry the client and the fabric, ANNY is the body, and RFD 1122 is the chain that puts a person into it. Two things follow and nothing else does.
People make things and wear them. A mesh someone made is worth nothing until it is rigged, textured and in the world, so TRELLIS.2 / Pixal3D takes 90 as the shortest path from an image to a wearable asset and SkinTokens 85, because an unrigged mesh cannot be worn at all. VoxHammer takes 60 for editing what exists.
People are present to each other, and keypoints are how. rf-detr is markerless motion capture: a camera, a body, and the pose that carries one person’s movement onto another’s avatar. That is embodiment rather than a step toward it, so it takes 85, and Kimodo 80 for moving a body without a camera in front of it. MuJoCo MJX takes 70, because a world that responds is the difference between a room and a backdrop.
EditScore takes 70, and an earlier revision had it at 10. That scored it as a thing nobody wears or sees, which is true and beside the point: it is a dependency of every making flow above it. A mesh generated, edited or restyled is a proposal, and something has to say whether the proposal is good before it reaches a wardrobe. Without a scorer the loops do not converge, so what people make is not made well. It is invisible and load-bearing at once, and 10 recorded only the first half.
What each row is for, and what its score rests on
The table above ranks. This says what each row would actually do in a chain and where its numbers came from, because a rank read without either invites the reader to supply their own.
model role in a chain basis
rf-detr keypoint live markerless pose, the measured,
capture end of every making flow 40.85 M params
cyclegan_style_transfer style transfer, 2D and 3D views read only
OmniGen2 text-guided 2D edit and corpus measured,
restyle 7.81 B params
EditScore, Qwen3-VL-8B the gate: scores a proposal inferred from
before it is accepted code, 6.75 GiB
Kimodo pose without a camera RFD 1026 estimate
MoGe depth and intrinsics from one read only, a
photograph, the ingest bootstrap Pixal3D dependency
SkinTokens rigging and skinning, mesh to RFD 1026 estimate
wearable asset
MuJoCo MJX body dynamics and collision, runtime only,
the presence goal not a network
Mitsuba 3 shading forward shading, G-buffer and measured, 14.1 ns
MToon per pixel a pixel
TRELLIS.2 / Pixal3D image to sparse 3D structure, measured,
the 3D backbone 24.04 GB weights
VoxHammer edit an existing asset in latent read only, then
space its ViT measured
Qwen3-ASR-1.7B transcription and alignment, the measured, 2.04 B,
ear of every making flow enc 1024 x 24
Qwen3-TTS CustomVoice a cloned voice, host-side measured, 1.92 B,
12 Hz codec
LaMa, AnimeMangaInpaint fills what a mask removed, so a measured, MIT,
layer comes out whole rfftn refusedTwo rows are not networks and the basis column says so. MuJoCo MJX is a physics engine and Mitsuba 3 is a renderer. Both were measured at runtime and neither has weights or a graph, so their shape and reference scores answer a question the models were asked and these two cannot be. RFD 1167 puts them off the ladder for the same reason.
Mitsuba’s 14.1 ns is the forward shading pass alone – fixed-dimension per-pixel G-buffer and MToon arithmetic. Backward ray tracing and BVH traversal are excluded, and that exclusion is the whole reason the row scores 75 on shape: what remains is dense arithmetic at a fixed size, and pointer-chasing traversal is not.
The citations, and the eight this workspace had to correct
references/ holds a CITATION.cff for each row. They exist because a candidate table that names twelve projects and credits none of them is a provenance failure of the kind the blocklist exists to catch.
A first draft of these files was supplied and eight of the twelve attributed work to the wrong party, so every entry was checked against the manifest’s pinned parents and the LICENSE in each checkout rather than accepted:
row claimed actually
rf-detr keypoint Weftspun Team Roboflow
OmniGen2 VectorSpaceLab/OmniGen .../OmniGen2
EditScore QwenLM/Qwen-VL VectorSpaceLab/EditScore,
a LoRA over Qwen3-VL
Kimodo Weftspun Team a weftspun wrapper only;
upstream unverified
MoGe RuidaZhang/MoGe microsoft/MoGe
SkinTokens Weftspun Team a weftspun wrapper only
See-Through Weftspun Team shitagaki-lab/see-through
TRELLIS.2 / Pixal3D microsoft/TRELLIS microsoft/TRELLIS.2 and
TencentARC/Pixal3D, two
upstreams in one row
VoxHammer Weftspun Team Nelipot-Lee/VoxHammerFive of those credited someone else’s model to this workspace, which is the direction of error that matters. Licenses are read from the LICENSE file in each checkout and not inferred: MIT for MoGe, TRELLIS.2, Pixal3D and VoxHammer; Apache-2.0 for OmniGen2, EditScore, See-Through and MuJoCo. Kimodo, SkinTokens and rf-detr-cpp carry no LICENSE file, so none is claimed for them.
Where a row is a weftspun HTTP wrapper around a model held elsewhere, the file says so and tells the reader not to cite it as the model.
See-Through leaves the ranking and stays as a taxonomy
Decision: Marigold is dropped and See-Through is kept as a reference for layer taxonomy rather than as a model. It is no longer an acceleration candidate, so it is no longer a row, and the table above holds eleven.
The reason is that nothing in its weight set can be used. The repository is Apache-2.0 and that covers the code; every checkpoint the inference scripts actually load is hosted separately and states no licence at all, which is not permissive – absent a grant the default is all rights reserved:
layerdifforg/seethroughv0.0.1_marigold none stated (moved from 24yearsold/)
24yearsold/l2d_sam_iter2 none stated
24yearsold/seethroughv0.0.2_layerdiff3d_nf4 none stated
24yearsold/metricdepth3d_tmp HTTP 401, unreadableThe depth checkpoint is additionally a fine-tune of prs-eth/marigold-depth-v1-1, which is CreativeML Open RAIL++-M over stable-diffusion-2. A fine-tune does not reset that licence. RAIL requires its use restrictions to travel into derivatives, which is the mechanism the OpenRAIL-M blocklist row already names, and nobody downstream can relicense away restrictions they do not hold. Shipping alongside Apache-2.0 code does not extend that licence to weights; a project meaning to do so says so, as NVIDIA does for Kimodo.
Replacing the depth stage with MoGe was considered and is not a substitution. MoGe is MIT and already in the manifest, but Marigold here is a latent diffusion estimator – VAE, UNet and CLIP, with cvt_marigold2d_to_3d.py around it – while MoGe is a feed-forward ViT emitting point maps. That is a rewrite of the stage, and the fine-tune exists because See-Through works on illustration where MoGe is trained on photographs. Whether MoGe holds up there is unmeasured.
What is kept costs nothing and is the useful part. The layer taxonomy is a vocabulary, and common/live2d/scrap_model.py carries three revisions of it:
V1 20 parts hair, skin, beard, and single-part eyes
V2 20 parts drops skin and beard, adds objects
V3 23 parts front hair, back hair, headwear, face, irides,
eyebrow, eyewhite, eyelash, eyewear, ears, earwear,
nose, mouth, neck, neckwear, topwear, handwear,
bottomwear, legwear, footwear, tail, wings, objectsThe V2-to-V3 change is the one worth having: hair splits into front and back, and the eye splits into irides, eyewhite, eyelash and eyebrow. That split is exactly what CLAUDE.md points at when it says a photograph has no ground-truth front hair / back hair, and it is why the blinded COCO holdout cannot validate this task whatever else it is good for.
Removing the row moved two others, and that is a property of the method rather than a mistake. With See-Through present, Mitsuba 3 took seat 9 and TRELLIS.2 / Pixal3D seat 11; without it they swap to 9 and 10. STAR’s runoff is not independent of irrelevant alternatives, so dropping a candidate can reorder the ones that remain. The recorded order is the one the eleven-row election produces, and check_rfd1166_rank.py recomputes it against starvote rather than trusting the arithmetic above.
A row is not a model, and that is systemic
Every row here names a project. Not one of them is a single network, and the composite is where both the licence and the accelerable part live. Scoring and citing at row granularity hides both.
row the part that would compile the part that binds the licence
rf-detr keypoint a patch-12 DINOv2 backbone Roboflow, Apache-2.0
EditScore Qwen3-VL's stock ViT a LoRA over someone else's
base model
MoGe DINOv2 ViT-B/14, four microsoft/MoGe, MIT
intermediate layers
VoxHammer stock dinov2_vitl14_reg Nelipot-Lee, MIT, over a
at 518 square TRELLIS backbone
Kimodo not examined code Apache-2.0, weights
NVIDIA Open Model, one
variant NVIDIA R&D
See-Through not examined code Apache-2.0, depth
OpenRAIL++-M, one checkpoint
unreadable
TRELLIS.2 / Pixal3D the SS stage alone two upstreams, Microsoft
and TencentARCFour of the twelve rows compile because of DINOv2, and none of them compile the same DINOv2. rf-detr runs a custom patch-12 variant, MoGe a ViT-B/14 taking intermediate layers, VoxHammer a stock ViT-L/14 with registers at 518 square, EditScore the ViT inside Qwen3-VL. One HEF does not serve them, and the tempting conclusion that it might is the reason to write the sub-model down rather than the row.
Kimodo’s Apache-2.0 badge covers the code and not the weights. Kimodo-SOMA-* and Kimodo-G1-* are under the NVIDIA Open Model Licence, which permits commercial use; Kimodo-SMPLX-RP-v1 is under NVIDIA’s Internal Scientific Research and Development licence, which does not. Reading the badge and stopping there answers for neither.
The restrictive variant turns out not to constrain this row, and the reason is the rig rather than the licence. Motion travels SOMA-X to ANNY to Godot Humanoid to VRM, so no SMPL-X checkpoint is on the route – and SMPL-X is blocklisted besides. Kimodo-SOMA-* matches the rig and is the shippable variant, which is the same conclusion logbook-rfd1016-model-repos.md reached when it warned against swapping checkpoints without re-checking RFD 1028.
Retraining Kimodo to get an Apache-2.0 or MIT artifact was considered and is not available, and the reason is the data rather than the model.
The NVIDIA Open Model Licence does not force it. It permits commercial use, permits derivative models, and states that the derivative is owned by whoever made it – it is not share-alike, so nothing propagates the way OpenRAIL-M’s restrictions do. What it does attach is attribution, a guardrail-circumvention clause that terminates the grant, a patent-litigation trigger, and NVIDIA’s Trustworthy AI terms. Those are conditions on use, not a bar to shipping, and logbook-rfd1016-model-repos.md already recorded Kimodo-SOMA as the shippable variant on that basis.
Retraining would buy an artifact with no third-party terms at all. It cannot be done on Kimodo’s own corpus. BONES-SEED sits behind an acceptance gate – 288 hours, 142,220 annotated animations, 522 performers, under a custom bones-seed-license whose terms are not readable without accepting them first. That is the condition CLAUDE.md names in as many words: a set behind a registration form is not licence-clean, because terms that cannot be read cannot be gated on. It is the same test that blocklists 24yearsold/metricdepth3d_tmp two sections above.
So the licence-clean motion pool is what remains after the blocklist takes CMU on provenance, Mixamo on licensing and posemaniacs on scraping: ANNY/SOMA’s own pose library, constructed synthetic rendered from assets held here, or a CC-BY-4.0 capture set with citation metadata. A retrain is a data programme before it is a training run, and pretending otherwise would put the effort in the wrong place.
The recommendation is therefore to ship Kimodo-SOMA-* under its licence and treat the retrain as a separate decision about corpus independence, not about whether this row can be used.
See-Through is the sharpest case, and it reaches the blocklist. The repository is Apache-2.0 and pulls two checkpoints that are not:
prs-eth/marigold-depth-v1-1is CreativeML Open RAIL++-M, fine-tuned fromstable-diffusion-2. CLAUDE.md blocklists OpenRAIL-M as a generator because its use-restrictions propagate into anything trained on the output, and exempts passthrough use. Which of the two this is depends on whether See-Through’s output enters a corpus – andseethrough PSDsis already a blocklist row for secondary generation. Those are the same fact reached from two directions.24yearsold/metricdepth3d_tmpreturns HTTP 401. No readable licence, no model card, and a name ending_tmp. CLAUDE.md is explicit that terms which cannot be read without accepting them cannot be gated on, so this is not licence-clean.
Neither was visible from the row. Both were found by reading what the code loads, which is the only method that works here.
So the rule this section exists to state: a candidate is cleared, or scored, or cited, at the granularity of the checkpoints it loads. The row is a name for a bundle, and a bundle has no licence and no operator census of its own.
Three rows added, and where their numbers come from
RFDs 1168 and 1169 turned up candidates the ranking had no row for. They are scored on the same eight dimensions, and because these scores were assigned here rather than inherited, the basis for each is stated.
Qwen3-ASR-1.7B enters at seat 2, behind only rf-detr.
fit 95 2.04 B, 2.04 GB at eight bits against 8 GB. Measured. It
never meets RFD 1128's four-bit question at all.
shp 85 the encoder is fixed-shape and every operator it needs is
in DEVICE_OPS; `PROJECTOR_TYPE_QWEN3A`'s conv2d front end
says the log-mel is image-shaped. Decode is on the host.
ref 45 rf-detr's device half is the precedent for the shape, and
there is no Hailo audio reference anywhere. Split the
difference rather than pretend either way.
clr 80 Apache-2.0, ungated, newer than our base. The only
obstacle is widening our own `clip_init_hailo`.
val 80 it is the only row that adds a modality
adp 75 1.7 B is trainable on hardware this workspace has
ask 95 downloadable and runnable today
wnt 85 sound was the reason Gemma 4 was wanted at allQwen3-TTS CustomVoice at seat 4 scores shape 25 and reference 15 because the autoregressive stage does not compile, and clear 80 because nothing else about it is in the way. It is a host-side row that earns its place on fit, ask and wanted rather than on accelerability, which is the honest shape of it.
LaMa at seat 5 scores shape 5 and reference 5 on a measured refusal, not an estimate: ffc.py calls torch.fft.rfftn and no Fourier operator appears in DEVICE_OPS. clear is 85, the highest in the table, because MIT weights over Apache-2.0 code with no gate is as clean as a row gets here.
Three of the five top seats are now host-side rows. That is a real change to what this table says. It ranks what is worth doing, and what is worth doing has stopped being the same question as what compiles – Qwen3-ASR’s encoder is the only one of the three new rows with a device half at all.
These scores are mine and are the weakest part of the table. Every other row was scored by somebody deciding; these were derived from measurements taken the same day. value and wanted in particular are judgements about the product that a measurement cannot settle.
How much of this needs original research, and the answer is almost none
Worth asking before anyone’s time is committed, because the eight dimensions do not separate hard from unknown. Sorting the field by what actually blocks each row:
blocker rows
hardware, nothing else rf-detr
wiring or an export nobody has run VoxHammer's DINOv2 stage,
MoGe, EditScore, OmniGen2,
CycleGAN, Kimodo, SkinTokens
a corpus RFD 1168's segmenter
a compiler refusal not yet understood TRELLIS.2 / Pixal3D
off this ladder MuJoCo MJX, Mitsuba 3Seven of the eleven are blocked on work, not on knowledge. Nothing in them asks a question the field has not answered. VoxHammer’s stage is a stock dinov2_vitl14_reg and MoGe’s is a stock ViT-B/14 – both exported today, both clean against DEVICE_OPS, both needing a DFC run that costs an afternoon. Kimodo and SkinTokens are NotImplementedError in a weftspun wrapper with a licence-clean upstream sitting behind it. EditScore needs an encoder built with its merger LoRA merged, which RFD 1157 already specifies. CycleGAN needs its checkpoint fetched.
rf-detr is blocked on a card and nothing else. 32.5 GiB against the desk’s 24 is a purchase or a rental, and RFD 1165 carries the number. The model translates, the calibration set exists, the compile script exists. That is the whole gap between rung 3 and rung 5.
One row is genuinely open. Pixal3D dies in _add_input_layers on input rank, four times, and whether that is a graph restructuring or a real limit of the parser has not been established. It is the only place in this table where somebody would have to find something out rather than build something.
One is a data programme wearing a model’s clothes. RFD 1168’s segmenter needs VALID_BODY_PARTS_V3 labels that do not exist here. The technique is ordinary supervised segmentation; the corpus is the work, and RFD 1168 records that corrupt-clean render pairs generate it from assets already held. Kimodo’s retrain is the same shape and is worse off, because BONES-SEED is gated.
So the honest summary is that this workspace is short of compute and wiring, not of ideas. That is the more actionable finding, and it is the opposite of what a ranking with a reference column tends to suggest – low reference scores read as “nobody knows how”, when for most of these rows they mean “nobody here has run it yet”.
What this ranking does not tell you
It ranks units of work, not deliverable capabilities, and two things follow that a ladder would otherwise walk into.
A chain delivers when every stage is placed, not when its best stage is. Accelerating one row leaves the others on the host, which is often fine and sometimes the whole point – rf-detr is worth doing alone. But no row’s rank is a claim that finishing it finishes anything.
Two rows interact, and the table shows them as separate jobs. TRELLIS.2 / Pixal3D sits at 11 and VoxHammer at 12 as though each could be taken on alone. An earlier revision of this section said they were mutually exclusive, on the reasoning that a compiled graph has no Python forward for VoxHammer to replace. That was asserted from a general principle and the code does not support it.
The patched attention blends a cached key against a live one:
k = k * kv_mask + ss_kv[...].cuda() * (1 - kv_mask)Mul, Sub, Mul, Add, all of them already inside DEVICE_OPS. The arithmetic compiles. What is Python is the dict lookup, and a compiled graph would take those cached tensors as inputs rather than fetching them.
What actually bites is the interface, not the patching. The cache is keyed by timestep, order, position, layer and type, so the tensors crossing the boundary scale with all five, and the cache is state held between the inversion pass and the edit pass. On a part that pays 2.18 ms a dispatch, marshalling that per timestep is the same arithmetic that rules out a per-node ggml backend.
So the two rows are not a fork. They are one job that is larger than either row suggests, and nothing here has measured how much larger.
The critical path is a chain, not a set of loops
value was first scored per invocation, which treated the models as independent candidates. They are stages of one shape:
2D edit CycleGAN -> OmniGen2 -> EditScore
3D edit CycleGAN -> VoxHammer over TRELLIS.2/Pixal3D -> EditScore
fit rf-detr -> ANNY fit -> EditScore
+ soma_refereeOmniGen2 appears in one chain and not the other, which is a decision rather than an omission. See below.
The two are siblings here and the corpus already orders them. pose-consensus/README.md records the inversion: instead of estimating poses from images, ANNY originates the pose, renders it, and a diffusion model stylizes the render with the geometry pinned, so the label is true by construction. That inversion ended the consensus panel outright – you do not convene three estimators to adjudicate a pose you authored.
Run that way the chain is one, not two:
originate 3D -> render views -> style views -> edit -> scoreand 2D is an operation on views rather than a stage before 3D. Three things follow.
The camera question disappears. OmniGen2’s 0.04 obedience slope only matters if the camera is asked for. Rendering from sphere_hammersley_sequence gives it exactly, and the measurement below stops being a problem to work around.
The labels land on the right side of CLAUDE.md’s synthetic rule. A render is constructed, deterministic from assets held here, with labels true by construction. An image generated first and reconstructed after is generated, and carries all five conditions.
rf-detr changes job without changing score. Estimating an unknown pose and confirming an authored one survived a restyle are different questions, and test_pose_survives_restyle.py asks the second. Its wanted 85 covers the live case, a person in front of a camera, which the inversion does not touch.
The 3D chain does not go through the 2D one. Pixal3D’s run() takes a single Image.Image, so reading it as an image flow suggests 2D must produce that image. It need not. get_cond one layer down takes list[Image.Image], and VoxHammer already drives that path: extract_feature.py renders 150 views, runs DINOv2 over them, projects voxel centres into each view and averages the patch tokens, reproducing the encoder input for an asset that already exists.
So the views come from rendering an asset rather than generating an image, which is also the cleaner side of CLAUDE.md’s synthetic line: constructed rather than generated. sphere_hammersley_sequence is the mandated sequence and Mitsuba renders a view in 1.79 ms.
Go through VoxHammer rather than calling the multi-view path directly. Both reach the same samplers, and only one of them is already written, tested against a real asset and handling the inversion the edit needs.
Three consequences, and each moved a row.
EditScore terminates every chain. Nothing reaches a wardrobe without passing it, so a gain there lands three times. Its wanted is 70 for being the dependency of making it does not itself perform.
CycleGAN styles both chains, and style is not 2D-only. A 3D asset is restyled through its views: render, restyle the views, ingest the styled views. It is the same stage doing the same work in both chains rather than a 2D step the 3D chain routes around, so value 70 and wanted 60. It holds second on a runoff it drew four-all, seated by score and the narrowest result in the table.
The hazard in that is already guarded. test_pose_survives_restyle.py restyles and re-detects to measure joint drift, and verify_restyle.py scores silhouette against the render’s own alpha with a shifted-reference control that must fail. A restyle that moves the body is the failure both exist to catch, and styling views before reconstruction is exactly where it would happen.
Mitsuba is the 3D chain’s view source, not a bystander, which takes value from 30 to 55 and wanted from 25 to 45.
OmniGen2 is 2D only, and the measurement is why
The 3D chain could route through it and does not. Camera control is what it would be for, and write_omnigen2_jsonl.py records what that costs: asked for eight azimuths in plain language, the recovered azimuth tracked the request with a slope of 0.04, where 1.00 is obedience and 0.00 is a body that never turned. Six of eight views came back between -1 and -11 degrees whatever was asked. The trained adapter reached 0.099.
The 3D chain does not ask. It renders the view it wants from sphere_hammersley_sequence, exactly, in 1.79 ms. Paying 133 s for a slope of 0.04 when a camera transform is exact and free is the trade this records, and it is why the 3D chain takes rendered views instead.
What OmniGen2 keeps is the 2D chain, and nothing else here can do it. CycleGAN is style transfer, one learned mapping. OmniGen2 is instructed editing: change this specific thing, in words. Those are different capabilities and the second is what “moddable” means. It is also anny-camera-lora’s base, so dropping it orphans RFD 1141’s published adapter, and it generates the corpus restyles.
Its value of 95 is the highest here because it is 133 s of a 163 s round, so removing it takes out the largest cost in the loop. value measures what accelerating a stage buys, not whether the stage should exist, and on OmniGen2 those two point in opposite directions.
VoxHammer has a device half, and it took three tries to see it
It holds no weights: no nn.Module, no nn.Parameter and no checkpoint in 3-interactor/voxhammer-upstream. An early revision scored it n/a on that basis, which answered the wrong question – weighing nothing does not put it outside a ranking of work.
shape was then 0, on the reasoning that VoxHammer replaces forwards at runtime and a compiled graph has none to replace. That was wrong, and defended twice before anybody read the stage it describes.
extract_feature.py loads dinov2_vitl14_reg and runs it at 518 square, n_patch = 518 // 14, under the same ImageNet constants compile_hef.py folds into an input layer, 150 times per asset. A stock ViT at fixed resolution is the same family as rf-detr’s device half, which translated at 825 nodes with the allowlist holding. reference is 55 for that precedent rather than 0.
What does not compile is separable and was never the whole model: F.grid_sample projecting voxel centres into views, and the N x M coordinate match with its argmax and scatter. Both are RFD 1131’s refused family and both sit either side of the encoder, which is the cut rf-detr already uses. So shape is 50.
fit 30 and clear 15 are the model’s own. ask 30 is not: upstream runs, and the two weftspun wrappers raise NotImplementedError outside stub mode, a wiring gap here. value 60 and wanted 80 follow from the chain, where it is two stages and the 3D route has no other way in.
It still ranks last, now for an honest reason: four columns are low, and one strong column does not carry a row through a runoff. RFD 1167 puts it on rung 0 with a rung-1 stage waiting – the shortest export left in the field.
Five models the earlier revision omitted
cyclegan_style_transfer is the serious omission. It is Loop 3, called at localhost:8000 by 3-stylized-to-omnigen2.livemd before OmniGen2 runs, so the ranking claimed to cover the candidates while skipping a model already inside the pipeline the accelerator serves. It is also the accelerator’s home ground: fixed-resolution image-to-image convolution, no autoregression, no sparse indexing.
MoGe recovers point maps, depth, normals and camera FOV from one image, and depth estimation is a category the Hailo zoo already ships for HAILO10H, which is what its reference 55 records.
MuJoCo MJX serves presence, and was twice mis-filed. First as tooling, then as interactability at the edge of the product. Physics is embodiment: a body that collides, stands on ground and responds is present, and one that floats and clips is not. Its wanted is 80, with rf-detr and Kimodo.
mjx/mujoco/mjx is MuJoCo reimplemented in JAX: a differentiable physics graph, made fixed-shape on purpose through fixed-size contact buffers, which is why adapt is 90 – gradients run through the simulator itself.
Its blocker is the interchange. JAX lowers to StableHLO through XLA and the Dataflow Compiler consumes TensorFlow, TFLite and ONNX, so there is no path today. That is a different failure from GGUF’s: GGUF carries no graph at all and nothing can convert it, while StableHLO is a graph in the wrong dialect. Blocked on format, not structurally, and clear 15 says which.
Mitsuba 3 is scored as its shading pass, not as the renderer. It is a pixi dependency of anny-render-corpus at 3.9.1 rather than a manifest project, and the row covers only the half that could go.
Ray tracing cannot: BVH traversal is pointer chasing with data-dependent branch depth, which is what a dataflow part is least able to express. Shading over a G-buffer is the opposite – per-pixel arithmetic at fixed dimensions – and the corpus already separates them, at 48.1 ns a pixel for the G-buffer against 14.1 for MToon shading.
Two things hold it at 220 anyway. Dr.Jit compiles to LLVM IR and PTX, which is machine code rather than a portable graph, so it is a level below even MJX’s wrong-dialect problem, and whether the Slang shading path can be expressed as ONNX at all is unverified. And the ceiling is small: the same measurements put intersection at 77 per cent of the combined cost, so the whole prize is the other 23 of a pass already made 1,272 times faster.
Not models, correctly absent: tropes-removal-model is ste-enforcer, a prose linter under a misleading directory name; mujoco-riscv64 is an engine port; anny and soma-x are parametric bodies; pose-consensus is a solver; bumblebee is a framework.
What each row rests on
row basis
rf-detr MEASURED. 40.852 M parameters from the
checkpoint, 25.245 M in the device half.
Exported and translated.
OmniGen2 MEASURED. 15.87 + 15.02 + 0.34 GB on disk at
fp32, so 7.81 B. Trained here: anny-camera-lora,
200 steps at 256 square in 22 minutes.
Pixal3D MEASURED. Repo safetensors total 24.04 GB
against RFD 1026's estimated 24.05.
See-Through RFD 1026 ESTIMATE, and the four components it
covers are not recorded. Only `layerdiff` is in
this table; `marigold-depth`, `vae` and
`partseg` are separate repos.
TRELLIS.2 RFD 1026 estimate.
SkinTokens RFD 1026 estimate.
Kimodo RFD 1026 estimate.
EditScore INFERRED, and the inference was wrong once:
scoring assumed Qwen3-VL-4B and `weft_score.py`
loads the 8B.
CycleGAN UNEXAMINED. Scored from what it is, not from
anything run. `clear` 50 says so.The adapt column, and the one number that bounds it
It leans on RFD 1140: the desk trains an OmniGen2 LoRA at 256 square in 22 minutes and does not finish one step in twelve at 512. Pixal3D’s 20 and Gemma 4’s 15 come from that, and from GGUF carrying no graph to train against.
One kind of adaptation is out of reach entirely, at any size in this table. Quantization-aware fine-tuning exhausts the desk’s 24 GiB on rf-detr’s own device half – 25.245 M parameters, the smallest graph here – at the default batch and again at batch_size=1, epochs=1 over 64 frames, the directive verified in the loaded model script. RFD 1165 carries the retraction that batch size was ever the lever.
So adapt scores ordinary training and LoRA, which the desk does. It does not score QAT, which the desk does not do for anything, and a column that scored both would rate every row zero and say nothing. The distinction matters because RFD 1128’s four-bit question wants QAT: compression without fine-tuning is what optimization level 1 gives, and that is a different artifact rather than a slower one.