Logbook: RFD 1016 model-repo build-out
Source: RFD 1016’s inventory table (15 models). Per the user’s explicit override, one standalone repo per model — not RFD 1036’s “one repo, many folders” convention.
Scope decided
- 15 catalog entries total.
- 1 already satisfied elsewhere:
misamaru_seethrough— covered by the existinginteractor-seethrough-ggml/interactor-seethrough-torchrepos from earlier in this work. - 4 abandoned, no repo created, per each RFD’s own
State: abandonedfield (RFD 1064’s pivot away from scene/world reconstruction toward character concepts):worldmirror2_reconstruct,triposplat_image_to_splat,weftspun_image_to_world,lingbot_map_environment_scan. - 10 repos built — the rest of this log.
The 10 repos, in priority order
Priority = how ready each is to actually run for real, not the order built in. Ranked by: does it unblock other repos, is the license clean and confirmed, and is the upstream API already verified (vs. still an honest NotImplementedError).
| # | Repo | RFD | License | Why this rank |
|---|---|---|---|---|
| 1 | interactor-trellis2-image-to-textured-mesh | 1038 | MIT | Dependency root — 3 other repos build FROM weftspun/trellis2-base or share its weights. Blocks the most work if left undone. Real upstream API verified against microsoft/TRELLIS.2’s own README, _run_upstream() implemented for real (not a stub). |
| 2 | interactor-pixal3d-image-to-textured-mesh | 1040 | MIT | Ported verbatim from RFD 1040, which was already the fully-worked complete example for RFD 1036. Ship-ready once built and smoke-tested — no NotImplementedError anywhere. |
| 3 | interactor-qwen-image-edit | 1043 | Apache-2.0, independently checked | Clean license, standalone (no dependency), high catalog value (2D image edit, distinct capability from every mesh model). Only gap: the diffusers QwenImageEditPipeline call against the Q4_K_M GGUF isn’t verified yet. |
| 4 | interactor-skintokens-auto-rig | 1046 | MIT, confirmed against the real LICENSE file (RFD 1046 had it “review pending” — it isn’t ambiguous) |
Clean license, standalone, small model (1.0 GB bf16). Unblocks any downstream rigging pipeline. |
| 5 | interactor-trellis2-image-mesh-painting | 1039 | MIT | Depends on #1. Real upstream API verified against TRELLIS.2/app_texturing.py. Known gap: the xatlas UV pre-check RFD 1039 calls for isn’t implemented. |
| 6 | interactor-voxhammer-text-mesh-editing | 1047 | MIT, independently checked | Depends on #1. Composite (RFD 1037 taskweft HTN domain), domain.ex/problem.ex/plan.ex ported verbatim, dispatch order proven in stub mode. Gap: VoxHammer’s real inversion/edit/splice/decode calls not yet wired. |
| 7 | interactor-voxhammer-image-mesh-editing | 1048 | MIT, independently checked | Depends on #1 and shares #6’s domain.ex (RFD 1048 has no domain of its own — only problem.ex/plan.ex differ, mode.conditioning: "image" in place of "text"). Same gap as #6. |
| 8 | interactor-kimodo-text-to-motion | 1045 | Code Apache-2.0; weights per-checkpoint — resolved to Kimodo-SOMA specifically (NVIDIA Open Model License, commercial-friendly). RFD 1045 hadn’t recorded a weight license at all; do not swap to the Kimodo-SMPLX checkpoint without re-checking RFD 1028 — that variant is the more restrictive NVIDIA R&D Model License. | Small model, license now resolved to a shippable variant. Gaps: the sampler call, RFD 1007’s validation gate, and the retarget path are all unwired. Retarget path corrected mid-build per user input to go SOMA → ANNY → Godot Humanoid → VRM via meshula/LabRCSF’s joints.csv canonical-joint pivot table, not a direct name guess. |
| 9 | interactor-krea2-turbo-text-to-image | 1042 | Krea 2 Community License — revenue-gated: free commercial use only under $1M company-wide annual revenue and <50 seats; larger orgs need a separate enterprise license. Not Apache/MIT. Flagged for RFD 1028’s owner: this clears the bar for a small deployer, not for every possible customer. | Largest model in the catalog (33.8 GB bf16 → 9.30 GB Q4_K_M), most build complexity (4-part staged load), and the only license here that’s conditionally gated rather than clean or resolved. Also: the exact HF repo id/filenames for the Q4_K_M GGUF set are an unconfirmed guess. |
| 10 | interactor-p3sam-mesh-segmentation | 1041 | Tencent Community License Agreement (territory-restricted — excludes EU/UK/South Korea), not MIT as RFD 1041 states. Verified by reading the real LICENSE file at Tencent-Hunyuan/Hunyuan3D-Part directly; confirmed no separately-MIT standalone P3-SAM repo exists. |
Lowest priority: the license is a real, unresolved gate (territory exclusions are a hard block for some customers, not a formality), and _run_upstream() isn’t yet verified against P3-SAM’s real model.py. Needs RFD 1041’s owner to correct the license field before this can ship. |
Amended: two of those rankings did not survive being run
The table above ranked by readiness, and readiness was judged from licences and API surfaces rather than from execution. Two rows have since been executed. Both moved, and one moved off the list entirely.
#3 interactor-qwen-image-edit is blocklisted. Its stated gap was that the diffusers call “isn’t verified yet”. It is verified now and it fails. Qwen-Image-Edit-2511 is 20.43B parameters needing about 38 GB at bf16 against this desk’s 24 GB, so it runs quantised or not at all — and quantised weights do not produce corpus data, which is generated-synthetic condition 5. The quantised path is also broken rather than merely forbidden: at NF4 it peaks at 11.9 GiB and speckles every pixel, scoring 0.098 to 0.719 on silhouette agreement against a control of 0.222.
Three explanations were eliminated rather than assumed. Not the torch version, since the corruption interactor-pixal3d recorded under torch 2.4.1 reproduces unchanged on 2.11.0+cu128. Not the guidance, since diffusers silently ignores true_cfg_scale when no negative prompt is supplied and enabling it changed nothing. Not the input, since OmniGen2 edits the same render cleanly. 8-bit would have isolated the quantiser and cannot run here: step 3 of 30 at 4,925 s/it with 42 GB resident, 37 hours projected for one image.
The ranking’s own logic is what to correct, not just the row. “Clean licence, standalone, high catalog value” was true of this model and told us nothing about whether it works: Apache-2.0 in base and control alike, and unusable on the only hardware we have. A readiness table that never runs anything ranks paperwork.
#2 interactor-pixal3d-image-to-textured-mesh is not ship-ready. It is described as verbatim-complete with no NotImplementedError anywhere, which is true of the code and not of the image. Its requirements-hfdemo.txt pins a natten wheel whose kernels are sm_90 only — 182 cubins, all sm_90a, and no PTX at all, read off the wheel with cuobjdump — so on any non-Hopper rental it imports cleanly and then dies inside a diffusion step. vast.ai rents whatever is free. Triton also JIT-compiles a C shim on first launch, so a compiler and the python headers are runtime dependencies of an image that compiles nothing.
Separately, the base image that produced its two sample meshes is gone: weftspun-pixal3d:cdbb2bb was never pushed to a registry, its recipe was never tracked, and both Docker contexts now report zero images. What survives is a digest in a build log that fetches nothing.
#9 interactor-krea2-turbo-text-to-image is blocklisted, on the ground this table already named. The row called its licence “revenue-gated … clears the bar for a small deployer, not for every possible customer” and ranked it 9 rather than excluding it. That flag is now resolved: a use restriction whose satisfaction depends on who deploys the trained model cannot be gated on when the corpus is built, which is the same reasoning that blocks OpenRAIL-M as a generator.
A second reason arrived independently. The row’s plan was the Q4_K_M GGUF set, since 33.8 GB bf16 is what made quantisation necessary — and quantised weights do not produce corpus data. The shape that made the model affordable is the shape that disqualifies its output, so the plan on file could not have produced usable corpus even with the licence resolved.
#10 interactor-p3sam-mesh-segmentation is blocklisted. The row ranked it last because “the licence is a real, unresolved gate”, and it is now resolved as a block rather than as a correction to make. Tencent’s Community License Agreement excludes the EU, the UK and South Korea, so the constraint is on who may run the tool rather than on what its output may do — and that is invisible to every check this workspace has. RFD 1041’s own record still says MIT.
OmniGen2 is the replacement for the editing slot, and it needs no exception: 7.8B, Apache-2.0 in weights and code, 17.3 GiB at bf16, clean output on the same input that defeated Qwen. At NF4 it fits 8 GB — 4.33 GiB of weights, 6.72 GiB peak — which is the only figure here that speaks to the ASUS UGen300, though fitting the memory is not the same as compiling for the device.
License corrections made to the org’s own record
Two RFDs stated a license that didn’t match the real upstream when checked directly:
- RFD 1041 says P3-SAM is MIT. Real upstream (
Tencent-Hunyuan/Hunyuan3D-Part) is under Tencent’s own Community License Agreement, territory-restricted. Not propagated — the new repo’s README marks it “review pending” and documents the discrepancy. - RFD 1046 marks SkinTokens’ license “review pending”. Real upstream (
VAST-AI-Research/SkinTokens) has a plain MITLICENSEfile — not ambiguous. Resolved to MIT in the new repo, documented for RFD 1046’s owner to close out.
One RFD (1045, Kimodo) simply didn’t record a weight license at all — resolved by reading nv-tlabs/kimodo’s own README: code is Apache-2.0, but checkpoints are licensed per variant (Kimodo-SOMA/-G1: NVIDIA Open Model License; Kimodo-SMPLX: NVIDIA R&D Model License, more restrictive). The new repo pins to Kimodo-SOMA specifically and documents why.
Known gaps left honest, not papered over
Every repo above the pixal3d line (which is verbatim-complete) has at least one NotImplementedError in server.py for a step whose real upstream call wasn’t verified in this pass — each one names exactly what needs confirming and against which upstream repo. None of these were guessed and left silent; STUB mode proves the request/response shape and (for the two composite VoxHammer repos) the dispatch order, without needing the real model.