Repository navigation
Releases
v0.6.4 — the memory-guard release
The memory-guard release. A job that won't fit in your machine's memory is refused before it starts, instead of freezing Windows. FLUX.2 needs 28 GB less memory, Wan clips get a 2× upscale, and ComfyUI 0.45.2 is the reference build. Every modality and the image agent loop were smoke-tested live on 0.45.2.
Before you upgrade
- FLUX.2 loads the fp8 text encoder. Download
mistral_3_small_flux2_fp8.safetensorsfromComfy-Org/flux2-dev(split_files/text_encoders/) intomodels/text_encoders/, or renameclip_namein the FLUX.2 templates if you keep the bf16 one. - The default relight runs on
flux2-dev-nvfp4-mixed.safetensors(fromblack-forest-labs/FLUX.2-dev-NVFP4). On a card before the RTX 50 series, nvfp4 upcasts and runs slower, so pass--model flux2_dev_fp8mixed.safetensors. - Wan
--upscaleneeds ComfyUI 0.45 and the RelayVSR files fromComfy-Org/RelayVSR:relayvsr_x2_lightx2v_int8_convrotindiffusion_models/,relayvsr_flash_vae_x2_fp32invae/. - ComfyUI-LTXVideo moves to
220e1d4.git checkout 220e1d4, thenpip install -r requirements.txt(it addscolour-scienceandOpenImageIO). Upstream fixed STG itself, so drop the local patch if you applied it. - freecad-mcp moves to v0.1.30 (
ec8abf2). Re-copyaddon/FreeCADMCP/into FreeCAD'sMod/, then restart FreeCAD and Claude Code. - Jobs can now be refused. A job whose models need more memory than you have left is stopped with the numbers (
render refused: …).chimera doctorshows the room, andCHIMERA_SKIP_MEMORY_CHECK=1runs it anyway.
The memory guard
On the 62 GB reference box, an LTX-2.3 job beside other heavy work pushed Windows' commit charge to 122 GB of a 107.6 GB limit, and the desktop froze twice. --reserve-vram isn't honoured under ComfyUI's dynamic VRAM. So before queueing, Chimera now:
- adds up what the job needs: the model files the graph loads, found through ComfyUI's own model folders, plus 2 GB;
- compares it with what's left: Windows' commit room, or available memory on Linux, minus 10% of RAM kept for the desktop;
- unloads, then refuses: if the job is short, ComfyUI unloads earlier models and the check runs again; if it's still short, it refuses and nothing is queued.
It covers generate.py and the agent loops. With it on, the same LTX job ran with commit at 79 of 108 GB.
From the model bake-off (each switched only on a clear win)
- Wan
video --upscaleis RelayVSR 2× (768×512 → 1536×1024). It was an error on Wan. It's 3.4× sharper than bicubic, adds no flicker to speak of (within 5% of the source), and takes 9 s on an RTX 5090. - FLUX.2's fp8 text encoder (16.8 GB instead of 33 GB) gave the same images at the same seeds. With nvfp4, a FLUX.2 relight needs ~40 GB instead of ~68 GB.
- Relight on Qwen-Image-2.1:
--mode relight --model qwen_image_2.1_int8_convrot.safetensors. It's more faithful and keeps more detail in ~18 GB. FLUX.2 stays the default, since it handled a two-light brief that Qwen missed. - MiniMax Music 3 vs ACE-Step for sung lyrics: 1–1, so no change.
Sub-brands
--brand family/name groups related brands under brands/family/name/, each with its own brand.yaml, for example a club's trophies, medals and certificates. File names stay flat (family-name_<mode>_<seed>).
Fixes
- A timed-out render is cancelled in ComfyUI (by its own id), so it no longer holds up the renders queued behind it.
chimera doctorandchimera freereport what they couldn't read or unload. A silent Ollama is no longer "idle".- CAD code-gen ignores inline
<think>reasoning. - Windows device names (
CON,NUL, …) are refused as brand names.
1571 tests (1523 on Windows CI). Full list in the CHANGELOG.
v0.6.3 — the review release
The review release: a full bug and documentation sweep of v0.6.2. Every modality was smoke-tested live, and every fix comes with a test that failed first.
Before you upgrade
- Same-seed renders keep both now. A render whose name is taken (same brand, mode and seed) replaces the old file only when it's an identical re-render. Otherwise it's saved as
<name>_2,_3, …. If a script of yours expected the old file to be overwritten, look for the_2. - The mesh3d and CAD loops need Pillow:
pip install -e ".[images]". They now say so up front instead of failing every iteration. - Stricter inputs.
audio --from-video/--fpswithout--mode foleyis now an error. ACE-Step keys, BPM and lengths outside what ComfyUI accepts are refused before anything runs. A brand.yaml with an unknown top-level key, or apositionoutside top/bottom-left/right or center, now warns or errors. - Replays of older LTX renders made at a size that isn't a multiple of 32 report drift (720 → 704): LTX always rendered at the snapped size, and the sidecar now says so.
Safer GPU hand-off
- A local Ollama or ComfyUI that times out (rather than refusing the connection) no longer reads as "holding nothing". The loop refuses to load the judge or start a render instead, which closes the path to the October PC freeze.
- A model bigger than the card waits until ComfyUI has actually emptied.
chimera freewaits for Ollama to let go before freeing ComfyUI.- A remote Ollama is never unloaded.
A judge that can't be fooled as easily
- These no longer pass:
- a NOT-MET on the verdict line;
PASS: false;- a reply cut off at
max_tokens; - a model's inline
<think>reasoning.
- A part that fails the geometry checks never wins best-of over a clean one.
- The CAD loop's code generator now gets the judge's "avoid …" terms and reasons; before, it only heard "correct the previous attempt".
3D printing
- The print check counts only cavities sealed off from the outside, so open cups, vases and tubes read SOLID, while leaky TRELLIS.2 sheets stay HOLLOW.
--solidmemory and scaling:- it uses at most about 2 GB of RAM (a chunky 230 mm print needed about 27 GB);
- its gaps scale down with small prints instead of webbing them shut;
- it drops specks but keeps real second bodies, and says what it dropped.
- Clear errors for empty or open meshes.
--color projectreally projects the image now.- FreeCAD: OBJs import upright, and a script without
RESULTexports its finished objects, not the inputs of its booleans. - Orientation: STLs converted with
3d --format stlstand up in a slicer, and turntables keep a long model in frame.
Brand kits, replay and the CLI
- A
nullin brand.yaml means the field's default, and top-level fields are type-checked. --no-watermarkturns a brand's default watermark off for one render.- A rotated phone photo is sized the way ComfyUI loads it, so the watermark lands on the image.
- Replay notices an input file that changed since the render.
- A job cancelled in ComfyUI's queue fails in about 15 s instead of waiting up to an hour.
- Uploads never replace another run's input file.
- LTX sizes snap to multiples of 32.
Docs
- About 30 corrections. The biggest: the Z-Image base download instructions fetched the Turbo weights.
- Also corrected: the FLUX.2 repo name, the blender_mcp launcher (it needs
--with "mcp[cli]~=1.12"), several LTX node claims and FreeCAD's bundled Python (3.11).
Validated live on an RTX 5090:
- Generation: ComfyUI 0.41.1, Blender 5.2.2 and FreeCAD 1.1.4 across image, video, audio, 3D, render, CAD,
--solidand replay. - Loops: a 2-iteration image loop and a 2-iteration CAD loop on Ollama 0.40.5 with the 27B judge.
1475 GPU-free tests · ruff clean · CI on Linux and Windows.
Full details are in the CHANGELOG. Every change since v0.6.2: v0.6.2...v0.6.3.
🤖 Generated with Claude Code
v0.6.2 — the print release
The print release: AI 3D models come out as one solid, printable STL in millimetres, and every finished STL is checked before it reaches a slicer.
Before you upgrade
- For 3D printing, install the new extra:
pip install -e ".[print]"(numpy, scipy, scikit-image). Everything else works without it;chimera doctorsays whether it's there. --mode finishSTLs are now in millimetres. A--scale-mmmodel, or any GLB/FBX, used to load 1000× too small in a slicer. If you've been scaling those up by 1000 in the slicer, stop. An unscaled STL/OBJ input, such as a FreeCAD part, keeps its own numbers.
render --mode finish --solid
TRELLIS.2 and Pixal3D don't output solids. Each mesh is a paper-thin, double-sided sheet with layers stacked inside (a shirt over the body), so the body is an empty cavity, and --watertight sealed that into a hollow shell. --solid (with --scale-mm) rebuilds it as one solid:
chimera render --mode finish --from figure.glb --solid --scale-mm 228.6 --formats stl,glb- Blender writes the mesh as a raw STL in mm at the print size, seams welded and small holes capped.
- The host finds the outside (closing leaks, with the radius picked from the jump in solid volume when the body cavity seals), keeps only the outer skin, and rebuilds one surface from it with an FFT Poisson solve.
- The finish keeps the largest piece and repairs decimation pinches.
It runs on the CPU: about 3 minutes and a few GB of RAM for a 230 mm figure. On an RTX 5090 box, a fresh TRELLIS.2 model at 200 mm went from HOLLOW (44 cm³ in 30 pieces, 6,478 open edges) to one 481 cm³ solid. Hunyuan3D 2.1 is already solid, so --watertight is still enough there.
STL print check
Every STL that --mode finish writes is checked for size in mm, closed and manifold edges, separate pieces, volume, and SOLID vs HOLLOW (rays cast through the model). The report prints in the console and goes into the sidecar as print_check; a HOLLOW result points you to --solid.
Fixes
- STL units, as above.
- The hero render after
--scale-mm. Scaling about the origin left the model floating, 0.39 m up for a GLB and 114 m up for a mm-scale mesh, away from the floor and the light rig: a dim or black render. The model now sits on the floor, centred, and the STL starts at z = 0.
Tips
- Judge a print with a low, raking light at full size; a shrunken overview hides surface defects.
- Generate 3D from images with a plain background, like a product shot. A busy scene behind the subject can come back as a plate behind the figure, which no clean-up removes.
Validated live on an RTX 5090 with Blender 5.2.2: --solid on four TRELLIS.2 and Hunyuan3D models, and the STL units for STL, --scale-mm and GLB inputs. 1230 GPU-free tests · ruff clean · CI on Linux and Windows.
Full details are in the CHANGELOG; every change since v0.6.1: v0.6.1...v0.6.2.
🤖 Generated with Claude Code
v0.6.1 — the upkeep release
The upkeep release: a maintenance routine that keeps watch on its own, and its first update window.
Before you upgrade
- Blender MCP add-on: reinstall it from the new pin,
lab/blender_mcp@dbbf836(addon/blender_mcp_addon/), then restart Claude Code so it launches the matching server from.mcp.json. - Upgrading Blender: the add-on and the "Allow Online Access" setting live in Blender's per-version settings, so a new Blender version starts without them. Choose Load Settings on its first-run screen, or reinstall the add-on. Headless
generate.py renderfinds the newest Blender install by itself. - FreeCAD on Windows: GHSA-2mfq-2h69-c873 (high, a DLL search-path issue) affects 1.1.4 and every earlier release. It is fixed in FreeCAD 26.3.0, which isn't released yet; the weekly report keeps it flagged until it is.
A maintenance routine (docs/MAINTENANCE.md)
-
Weekly report. It now also covers:
- the published security advisories that affect a pin, comfyui-mcp's pinned version, ComfyUI or a host app's reference version. It reads each advisory's version range, or goes by date when there is no range it can read;
- the host apps (Ollama, FreeCAD, Blender, Node.js, Python), compared with their latest releases and end-of-life dates.
The freecad-mcp advisory went unnoticed for three weeks because the report checked only commits. A pin held on purpose now shows as held, with the reason, and warns again only when upstream moves past what was re-audited.
-
Checklist issues open for the monthly update window (on the 1st) and the quarterly review (January, April, July, October).
-
CI also runs weekly, and on Python 3.13 and 3.14 as well as 3.12.
-
Dependabot also updates the pre-commit hooks, and ruff is pinned to one minor version everywhere.
Updates
- blender_mcp →
dbbf836, re-audited. Large responses such as screenshots and renders were cut off when the socket's send buffer filled, the bundled manual and API docs now cover Blender 5.2, and the MCP SDK is locked below 2. - Reference host apps: Blender 5.2.2 LTS (5.1 reached end of life in July) and Ollama 0.40.5.
- ComfyUI-LTXVideo stays at
15d09ab. Its newest commit was re-audited and is clean, but it changes none of the nodes Chimera uses, adds two dependencies, and still needs our STG patch. It moves once that fix lands upstream. - Reported upstream: the bugs v0.6.0 works around.
Validated live on an RTX 5090: the CAD self-correction loop (code-gen on Ollama 0.40.5, FreeCAD, a Blender 5.2.2 render and the judge; PASS on the first try), the image loop on Ollama 0.40.5, and the Blender 5.2.2 MCP bridge with the add-on from dbbf836. 1197 GPU-free tests · ruff clean · CI on Linux and Windows.
Full details are in the CHANGELOG; every change since v0.6.0: v0.6.0...v0.6.1.
🤖 Generated with Claude Code
v0.6.0 — the model refresh
The model-refresh release. The winners of the October 2026 quarterly model review are now the defaults for image, upscale, video, music and 3D, and the agent loops run on local Qwen models sized to your GPU. Replay now protects old renders from that switch. Also in this release: ComfyUI 0.40 and Ollama 0.40 compatibility, a fix for a GPU-memory clash between ComfyUI and Ollama that could freeze Windows, a freecad-mcp security fix, and a security review.
New defaults
| Was | Now | Why it won | |
|---|---|---|---|
| Image | Z-Image Turbo (brandless) / FLUX.2 (brands) | Qwen-Image-2.1 | Product shots become a reference edit that keeps the real product. About 6 s per 1024² image. |
--upscale |
4x-UltraSharp (ESRGAN) | SeedVR2 7B | Restores natural detail where ESRGAN invented texture. Adds about 10 s for 4096². |
| Video | LTX-2.3 | Wan 2.2, plus --camera moves |
Kept a locked camera locked (≤ 0.8% zoom, vs LTX's 9% push-in) and held identity. |
| Music | ACE-Step 1.5 | Stable Audio 3 Medium, plus audio --mode sfx |
Clear winner of the listening test. |
| 3D | Hunyuan3D 2.1 (shape only) | TRELLIS.2 (textured PBR) | Textured meshes, and it cuts the subject out of a busy photo. |
| Agent judge | the ComfyUI-QwenVL node | Qwen 3.8 27B on a local Ollama, with lighter Qwen3-VL judges on smaller GPUs | Best of six models on a new 34-case judge benchmark, at 1.5 s per judgment. Needs no node pack and no configuration. |
| CAD code-gen | none: you had to configure an LLM endpoint | Qwen 3.8 27B on a local Ollama | Built 5 of 6 test parts; Qwen 3.6 27B-coding built 4, Qwen 3.5 9B none. |
The earlier models are still one flag away: --model z_image_turbo_nvfp4.safetensors or flux2_dev_fp8mixed.safetensors (image), ltx23 (video with synced audio), acestep (music, which now sings --lyrics), hunyuan21 (3D), --upscale-model 4x-UltraSharp.pth, auto_generate.py --backend local, or a pin in the brand's brand.yaml.
Before you upgrade
- ComfyUI 0.40 or newer for the new defaults. They run on its core nodes, so no new node packs are needed. Older builds (0.24 and up) still run the earlier models through
--modelor a brand pin, plus--upscale-model 4x-UltraSharp.pthfor upscaling, and--mesh-model hunyuan21with a Z-Image or FLUX.2--modelfor the mesh3d loop.audio --mode sfxneeds 0.40. - Downloads: about 119 GB for all five defaults, or about 88 GB without the camera-move models. They come from ungated Comfy-Org repos, so no Hugging Face login is needed.
docs/SETUP.mdlists them, and each module'smodels.mdhas the exact files. - Existing brands:
- A brand that pins
defaults.modelkeeps that model. Brands made withnew-brandbefore v0.6.0 pin FLUX.2 (the old template wrote it), so they stay on FLUX.2. - A brand with no image pin moves from FLUX.2 to Qwen-Image-2.1.
- Video, music, 3D and
--upscalefollow the new defaults unless the brand setsvideo.model,audio.music_model,threed.modelordefaults.upscale_model.
- A brand that pins
- Video on Wan 2.2: clips are silent (add sound with
audio --mode foley, or use--model ltx23),video --upscaleworks only with--model ltx23(ask Wan for the size with--width/--height), and a brand'svideo:length, fps and audio settings apply to LTX only. Wan renders 81 frames at 16 fps unless you pass--length/--fps. - Old renders:
chimera replayre-runs on the model the render recorded and, with--comfy-output-dir, takes the render's standard name only when it is free, else<name>_replay, so it never replaces a file. It stops with a list of differences if the brand or a template changed since (pass--accept-driftto render anyway). To run older work on a newer model, usechimera replay <sidecar> --model <file>: only the prompt is compared, and a switch to another model family starts from that family's defaults. Sidecars from before v0.6.0 carry no settings record, so replay says what it can't check, and an ACE-Step one reports that the music graph's sampler shift moved from 1.73 to 3 (render it anyway with--accept-drift). A sidecar that recorded an absolute--from-image,--from-videoor--assetpath is refused: put the file in the brand folder (brandless: the folder you run from) and change the sidecar to the bare file name. - Replay renders on
http://127.0.0.1:8000unless you pass--comfy-url; it no longer uses the URL recorded in the sidecar. - Agent loops:
- With no flags or
CHIMERA_*variables,auto_generate.pynow runs on local Ollama models picked for your GPU (see the next section). Install Ollama and pull them first; with Ollama running,chimera doctornames your tier and prints theollama pullcommands. A run checks this before it renders anything and stops with the command if a model is missing. CHIMERA_*settings you exported for v0.5 still win over the tiers (unset them to get the new models), and aCHIMERA_LLM_*endpoint now drives the judge too, since--backend apiis the default. Nothing loads.env: export settings in your shell.--pipeline mesh3d --texture/--finalizenow need--mesh-model hunyuan21: TRELLIS.2 and Pixal3D meshes come out textured.
- With no flags or
- API keys:
OPENAI_API_KEYnow goes only tohttps://api.openai.com. If you pointed the agent loop at OpenRouter, Gemini, Groq or a LAN server with onlyOPENAI_API_KEYset, setCHIMERA_LLM_API_KEY(orCHIMERA_<ROLE>_API_KEY) instead. A run prints a note when a key it found isn't sent to your endpoint. - 3D and Blender exports:
render --mode finishnow defaults to--color auto, which keeps a textured mesh's materials (--color materialgives the old clay finish), and3d --format stl|objkeeps the GLB beside the STL/OBJ. - LTX-2.3 on ComfyUI 0.40 or newer needs the bundled LTXVideo patch (
modules/video/patches/). - FreeCAD users: update both halves of freecad-mcp: the pin in
.mcp.json, and the addon in FreeCAD'sMod/folder (copy the wholeaddon/FreeCADMCP/folder, since itspackage.xmlis now read at import, then restart FreeCAD). Also update FreeCAD itself to 1.1.4 or newer.
Agent loops: local models sized to your GPU
The defaults were chosen on a 32 GB RTX 5090, and most cards have less. So each agent-loop role you haven't configured (judge, prompt rewriter, CAD code-gen) falls back to a local Ollama model from a tier picked by your GPU's memory:
| Tier | GPU | Judge and rewriter | CAD code-gen | Pull |
|---|---|---|---|---|
| large | 32 GB+ | Qwen 3.8 27B (16.5 GB) | Qwen 3.8 27B | qwen3.8:27b |
| medium | 12–24 GB | Qwen3-VL 8B Instruct (6.6 GB) | Qwen 3.8 27B, partly on the CPU | qwen3-vl:8b-instruct, plus qwen3.8:27b for the CAD loop |
| small | under 12 GB | Qwen3-VL 4B Instruct (4.4 GB) | Qwen 3.8 27B, mostly on the CPU (slow) | qwen3-vl:4b-instruct, plus qwen3.8:27b for the CAD loop |
The sizes are at Ollama's 32k context; with room to spare, Ollama 0.40 gives a model more (the 8B took 16 GB on a free 32 GB card), and the GPU hand-off plans for that. A 24 GB card gets the medium tier: beside ComfyUI's image model it has no room for the 27B judge (about 22 GB with its context), so every judgment would wait on ComfyUI emptying itself. --llm-tier or CHIMERA_LLM_TIER sets the tier yourself. Per-role flags and CHIMERA_* variables still win, so any role can use a hosted model instead. Each run prints which local models it fell back to.
How they were picked, on the RTX 5090:
- Judge: a new benchmark of 34 hand-labeled Chimera outputs, 3 runs per model. Qwen 3.8 27B got 31.0 right (balanced accuracy 0.90, 1.5 s per judgment), Qwen3-VL 32B 30.0 (14.6 s), Qwen3-VL 8B 29.3 (0.8 s), Qwen 3.6 27B-coding 29.3, Qwen3-VL 4B 27.7 (0.6 s), Qwen3-VL 30B-A3B 27.3.
- CAD code-gen: six parts through the real loop. Qwen 3.8 27B built 5 (3 on the first try), Qwen 3.6 27B-coding 4, and Qwen 3.5 9B none of the three it tried. No lighter model held up, so every tier keeps the 27B.
Fixed: a self-correction run could freeze Windows
On Windows, ComfyUI and Ollama can't see each other's GPU memory: Ollama reported 30 GB free while ComfyUI held 15. After a day of renders ComfyUI held all but 0.1 GB of a 32 GB card, the loop's local judge loaded on top of it anyway, and the driver paged GPU memory to system RAM until the machine froze. The loops now keep the two apart:
- before a local Ollama model loads, they ask ComfyUI how much of the card it leaves free, counting (on Windows) what Ollama already holds. If it's too little, ComfyUI unloads its models first. If the room never comes, or ComfyUI stops answering after showing a full card, the model isn't loaded at all. The room asked for covers the full context Ollama may give the model: with the card free, Ollama 0.40 loaded the 8B judge at 16 GB, not the 6.6 GB it takes at 32k, and a first estimate of 9 GB would have let it load beside ComfyUI on a 24 GB card. Once ComfyUI has emptied, a model that fits at a 32k context loads anyway, since Ollama sizes its context to the room it sees, and the loop warns if Ollama ever takes more than planned (it reads Ollama's settings, such as
OLLAMA_KV_CACHE_TYPE, from its own environment: see SETUP); - Ollama models the run didn't load are unloaded before its first render (another app's included; set
CHIMERA_KEEP_LLM_LOADED=1only when Ollama runs on a second GPU), and the run's own before every render. Each unload waits until Ollama has let go, and a loop won't render beside a model that stays loaded; - the CAD loop unloads its code-gen model before Blender renders, and its judge and rewriter too.
The check reads a local CUDA card, so give a local Ollama or ComfyUI as localhost or 127.0.0.1. OLLAMA_HOST=0.0.0.0 now works too. chimera free and chimera doctor cover the same ground by hand. Don't run other heav...
v0.5.1 — blender_mcp bridge fix
Patch release: the blender_mcp bridge works again on fresh environments.
The 4309a396 pin bump gave uvx a fresh environment, which resolved blender_mcp's unbounded transitive mcp[cli] dependency to the new MCP SDK 2.x — upstream imports mcp.server.fastmcp, an SDK-1.x path that 2.x renamed, so the server died at import (MCP error -32000: Connection closed) before ever reaching Blender's socket. Old pins only appeared to work because their cached environments predated SDK 2.x.
.mcp.json now launches with uvx --with "mcp[cli]~=1.12" (≥ 1.12, < 2) — written ~= rather than <2 because the launch args transit cmd /c on Windows, where a literal < is a redirection character (#53, #54). The constraint comes off once upstream ships an SDK-2.x shim (freecad-mcp v0.1.22 already carries one and is unaffected).
Verified live post-fix: all three MCP bridges connect; Blender round-trips real scene data. Full details in the CHANGELOG.
🤖 Generated with Claude Code
v0.5.0 — the hardening release
The hardening release: a 53-finding, adversarially-verified whole-repo audit fixed end-to-end, plus a live re-validation of the entire stack on ComfyUI Desktop 0.33.3.
Highlights
- 21 verified bugs fixed, test-first — the watermark cluster (unreachable null-logo guard, wrong placement canvas, dead
opacityknob), the dead brand.yamlvideo:block, threeparse_verdictheuristic regressions, the winner-sidecar provenance mismatch, DCC edge paths (Blender runner parity, silent--formatsdrop, the black-render clay fallback), doctor/update-check honesty, ComfyUI client validation detail + history reuse, typed manifest sub-sections, and more. - Modularization —
generate.pyis now a 508-line parse+dispatch CLI (was 938): render/cad/finalize orchestration lives inbrandkit/dcc_cli.py, the sidecar/replay contract inbrandkit/replay.py, with shared stages (run_graph_to_file,eval_contact_sheet,even_azimuths) replacing hand-rolled copies and reserved-node-id guards on every graph splice. - Supply-chain upkeep — re-audited bumps: ComfyUI-LTXVideo →
15d09ab(also fixes the pack's load-failure on ComfyUI 0.33 core), freecad-mcp →3745ff9(v0.1.22, new read-onlyget_rpc_statushealth tool), blender_mcp →4309a396. The comfyui-mcp hold now covers 0.20.x–0.50.x — 0.50.x makesnpxinstalls auto-update-eligible, exactly the drift the.mcp.jsonopt-outs neutralize. - Resilient weekly checker — failure rows carry HTTP status, transient 5xx retried once; checker arms fully covered (94%).
- Reference build → ComfyUI Desktop 0.33.3, smoke-validated live across image, video, 3D, headless Blender render, headless FreeCAD cad, and the full mesh3d self-correction loop (Ollama Qwen3-VL judge).
566 GPU-free tests · ruff clean · cross-platform CI. Full details in the CHANGELOG.
🤖 Generated with Claude Code
v0.4.0 — FLUX.2 relight (--mode relight) + field-learnings docs
Adds image --mode relight — a FLUX.2 ReferenceLatent relight that changes a still's lighting while preserving its exact composition — plus brand-neutral, field-tested technique notes distilled from production use.
Added
image --mode relight(FLUX.2 ReferenceLatent). Relight/restyle a still keeping its exact composition:--asset <source> --subject "<relight prompt>". Encodes the source and rides it on the positive conditioning (full-denoise sample), which holds the layout that plain img2img drifts. FLUX.2-only — a Z-Image brand default transparently falls back toflux2_dev_fp8mixed.safetensors. New parity-guarded template + CLI + filler wiring, +4 tests. Live-validated on ComfyUI 0.26.2 (sunset + moonlight relights both preserved composition while transforming lighting).
Docs — field-tested technique notes (brand-neutral)
modules/image/recipes.md(new) — creator assets: platform emotes/stickers (GIF-not-APNG, background-cutout + dark-outline for light/dark legibility, matched-seed keyframes), tiered badge sets.- i2v motion-realism caveat (
modules/video/), 3D-print/casting prep (modules/threed/), and--texture(mesh3d Phase 4a) marked live-validated end-to-end (modules/agent/).
515 tests, ruff clean. See CHANGELOG.md for full detail.
🤖 Generated with Claude Code
v0.3.0 — Qwen3 agent layer, ComfyUI 0.26.2, CAD self-correction working end-to-end
The agent layer moves to Qwen3 (Ollama-unified) and the reference stack to ComfyUI 0.26.2 / FreeCAD 1.1.1 / Blender 5.1.2 — and the CAD self-correction loop now runs end-to-end (live-validated PASS, score 1.0). All four generative modalities re-smoked on the new stack.
Changed
- Agent layer → Qwen3 (Ollama-unified): judge = Qwen3-VL-8B-Instruct (
qwen3-vl:8b-instruct, non-thinking), text roles = Qwen3.6-27B. Qwen2.5 retired. Cloud (OpenAI/Anthropic/Gemini/OpenRouter) stays a drop-in. - ComfyUI reference build 0.24.1 → 0.26.2, full per-modality smoke (image · CAD · video · audio);
>=0.24.xfloor unchanged.
Fixed — CAD self-correction loop now passes end-to-end
Three masked bugs, each root-caused + fixed + tested:
- mm-scale CAD renders came out near-black → Blender light rig now scales to object size (+ clay for material-less imports).
- The judge read the 4-view contact sheet as N objects → count-agnostic contact-sheet preamble on the 3D rubric.
- The judge tag was a thinking model returning empty content → default to non-thinking
-instruct;LLMClient.chatnow errors clearly on empty content.
- FreeCAD 1.1.x cad-job compatibility (params via
\$CHIMERA_CAD_PARAMS).
511 tests, ruff clean. See CHANGELOG.md for full detail.
🤖 Generated with Claude Code
v0.2.2 - dependency re-audit + docs
Maintenance release — weekly dependency re-audit + docs.
Changed
- Dependency pins (re-audited per
docs/UPDATING.md): ComfyUI-LTXVideo229437c→4f45fd6, comfyui-mcp0.9.4→0.18.0, blender_mcp03004fd→98b0e49d, actions/checkoutv6→v7. comfyui-mcp 0.18.0's new Comfy Cloud / Civitai-MCP / auth-header / Claude-Agent-SDK features are all opt-in (env-gated or optional-deps omitted viaNPM_CONFIG_OMIT=optional).
Docs
- README restructured for a public audience + post-launch polish / licensing clarity + doc-accuracy fixes.
Full changelog: https://github.com/whartons/Chimera/blob/main/CHANGELOG.md