Skip to content

v0.6.0 — the model refresh

Choose a tag to compare

@whartons whartons released this 16 Oct 13:48
888cef7

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 --model or a brand pin, plus --upscale-model 4x-UltraSharp.pth for upscaling, and --mesh-model hunyuan21 with a Z-Image or FLUX.2 --model for the mesh3d loop. audio --mode sfx needs 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.md lists them, and each module's models.md has the exact files.
  • Existing brands:
    • A brand that pins defaults.model keeps that model. Brands made with new-brand before 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 --upscale follow the new defaults unless the brand sets video.model, audio.music_model, threed.model or defaults.upscale_model.
  • Video on Wan 2.2: clips are silent (add sound with audio --mode foley, or use --model ltx23), video --upscale works only with --model ltx23 (ask Wan for the size with --width / --height), and a brand's video: length, fps and audio settings apply to LTX only. Wan renders 81 frames at 16 fps unless you pass --length / --fps.
  • Old renders: chimera replay re-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-drift to render anyway). To run older work on a newer model, use chimera 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-video or --asset path 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://127.0.0.1:8000 unless you pass --comfy-url; it no longer uses the URL recorded in the sidecar.
  • Agent loops:
    • With no flags or CHIMERA_* variables, auto_generate.py now runs on local Ollama models picked for your GPU (see the next section). Install Ollama and pull them first; with Ollama running, chimera doctor names your tier and prints the ollama pull commands. 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 a CHIMERA_LLM_* endpoint now drives the judge too, since --backend api is the default. Nothing loads .env: export settings in your shell.
    • --pipeline mesh3d --texture / --finalize now need --mesh-model hunyuan21: TRELLIS.2 and Pixal3D meshes come out textured.
  • API keys: OPENAI_API_KEY now goes only to https://api.openai.com. If you pointed the agent loop at OpenRouter, Gemini, Groq or a LAN server with only OPENAI_API_KEY set, set CHIMERA_LLM_API_KEY (or CHIMERA_<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 finish now defaults to --color auto, which keeps a textured mesh's materials (--color material gives the old clay finish), and 3d --format stl|obj keeps 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's Mod/ folder (copy the whole addon/FreeCADMCP/ folder, since its package.xml is 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=1 only 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 heavy GPU jobs next to a loop.

Security

  • freecad-mcp → v0.1.27, fixing GHSA-v4pw-r6h3-v75q (HIGH). On v0.1.24 and earlier, while the addon's RPC server was running, any web page open in a local browser could make FreeCAD run arbitrary Python. Its new execute_code_headless tool requires per-call approval.
  • A security review of the repo and its full public history found no secrets. Fixed along the way:
    • LLM API keys go only to their own host: OPENAI_API_KEY / ANTHROPIC_API_KEY only to https://api.openai.com / https://api.anthropic.com, CHIMERA_LLM_API_KEY only to CHIMERA_LLM_BASE_URL's host when one is set (otherwise to each role's endpoint), and never onto an HTTP redirect.
    • Replay no longer trusts the sidecar it replays. A crafted sidecar could upload any local file to a URL of its choosing, and its mode or seed could name an output file outside outputs/. Replay now ignores the recorded ComfyUI URL, refuses absolute or .. input paths, accepts only the modes a modality has, whole-number seeds and media files as inputs, and every output name must be one plain file name, including the names ComfyUI reports. Sidecars now record input files by name (or relative path), never by absolute path.
    • FreeCAD and Blender jobs start without credential-named environment variables (*KEY*, *TOKEN*, *SECRET*, *PASS*, *AUTH*, proxies, webhooks, *_URL / *_URI connection strings, …), in every mode, so your own cad --mode script doesn't see them either, and only files in a FreeCAD job's own folder count as its output (an LLM-written script could print a forged result naming any file). The CAD loop runs LLM-written scripts, and their restrictions are a speed bump, not a sandbox.
    • 18 more comfyui-mcp tools need approval (downloads, ComfyUI-Manager security settings, extra model paths, custom-node scaffolding, paid API nodes, and more), and FreeCAD's reload_document, which discards unsaved edits, moved out of the read-only tier.
    • Pillow ≥ 12.3 for the images extra, CI actions pinned by commit, a detect-private-key pre-commit hook, and .gitignore rules for conventional secret files (.netrc, .npmrc, *.pem, credentials.json, …).
  • comfyui-mcp stays held at 0.18.0. Newer releases update themselves on every launch, which defeats pinning; the weekly re-scan is tracked in #56.

Also new

  • audio --mode sfx: short sound effects from a text brief.
  • audio --model acestep --lyrics "…": ACE-Step sings your words (section tags such as [Verse 1] / [Chorus]); the sidecar keeps them for replay. A lyric render is at least 40 s unless --duration says otherwise: in tests it never sang in 20 s, and at 40 s it sang on most seeds (5–6 of 7). Some seeds plan an instrumental anyway, so if a render comes out without vocals, try another --seed.
  • video --camera static|pan-*|zoom-*|rotate-* (Wan 2.2 Fun Camera) and --quality fast|full.
  • chimera free unloads local Ollama models, then frees ComfyUI's VRAM.
  • replay --model and replay --accept-drift.
  • render --mode finish --color auto keeps a textured mesh's materials and exports JPEG textures (the example brand's 77 MB GLB came out at 9.6 MB).
  • chimera doctor checks every modality's default model files.
  • Reasoning-effort controls for local and hosted LLMs, per role.

Fixes

  • The self-correction loop now converges with the local judge. When a subject's styling fought the brand (say, "a glossy candy-red toy rover" under a gunmetal brand), the Qwen3-VL-8B judge wrote contradictory fixes and each pass drifted further off-brand. The brand now wins:

    • the rubric judges the object itself;
    • fixes may not contradict each other, and a palette color the judge asks to avoid by name stays;
    • the FIX parser reads what small judges actually write, and keeps a phrase like "gaps between blades and hub" whole;
    • the loop drops the flagged styling from the subject but keeps the object's noun;
    • palette colors are named for the judge, with the hex kept.

    The README's headline GIF is that run, fully local: FAIL 0.40 → FAIL 0.49 → PASS 1.00 on one seed.

  • The judge could pass an image it had just failed on a checklist item ("NOT-MET: the head is completely absent ... Overall: PASS"). A criterion marked NOT-MET now fails the verdict, as the rubric always said, in any common form (1. Subject — NOT-MET, **Palette**: NOT_MET, a table cell, 1 NOT-MET in a summary, NOT‑MET with a typographic dash), while a line that only says nothing is unmet (No NOT-MET criteria., Criteria: 4 MET, 0 NOT-MET) keeps a pass, and Partially MET or a criterion label in the rubric's own negative phrasing (3. No missing or broken parts - NOT-MET) fails it. On a 32,190-reply grid of layouts, labels and spellings no false PASS or false FAIL is left. A last review rewrote 204 real judge replies 24,426 ways and closed three more ways a broken render passed: a criterion saying overall … would not pass taken for the verdict line (so a bare FAIL read as a PASS), a numbered criterion let off as a summary (2. NOT-MET: none (the top face is flat)), and softer wordings (Mostly MET, Barely met, MET (partially)). A PASS that admits an unmet criterion or is negated (Does not pass) fails too, and the rubric's instructions echoed back are ignored. No error is left on those 24,426 cases. Scores on other scales read right too: 4 out of 5 is 0.8 and 7 is 0.7, not 1.0.

  • A loop that repeated a seed (--seeds 7,7,7) reported a winner whose file held a later pass. Each pass now gets its own files, and a shorter --seeds list is topped up with fresh seeds.

  • Repeating an identical job failed with "cannot find the file". ComfyUI serves a job identical to an earlier one (same seed, prompt and model) from its cache and reports the earlier file, which Chimera had already moved. Each job now saves under a fresh file name.

  • Replay no longer overwrites the render it reproduces (including, on Windows, when the brand name was typed in a different case), keeps the recorded watermark decision, only frees the GPU once it knows it will render, and accepts every brand name new-brand creates. It also takes the other image and video types ComfyUI reads (.gif, .jfif, .avif, .heic, .m4v, …), names the field in a damaged sidecar instead of crashing, compares the words rather than the BPM/key suffix when music switches between Stable Audio 3 and ACE-Step, and replays foley without PyAV (the sidecar records the timing it used).

  • Video and product renders can start from the brand's own renders in outputs/images/ (the README's image → video → 3D chain said "not found"), and without Pillow, JPEG, WebP and GIF sizes are read from the file, so watermarked renders build the same graph either way.

  • Video lengths snap to the frame counts the models render, with a note when they change (ComfyUI floored any other length without a word).

  • Bad flag combinations are caught before the GPU is freed or anything is uploaded, and flags a mode would ignore (--upscale-model without --upscale, --bpm outside music) are errors. ComfyUI rejecting a job or timing out, an unknown brand, replaying a file that isn't a sidecar, finalize-texture without views, a Blender or FreeCAD job that lists a missing file, auto_generate with a bad --brand or --seeds, and a name the Windows console can't print are each one line instead of a traceback. A brand name that isn't a folder name is refused before anything renders.

  • CAD loop:

    • a part must come out as one solid; the judge had passed a fan rotor whose blades floated free of the hub;
    • scripts that import App work on FreeCAD 1.1;
    • names like socket_depth, or a shape named socket, no longer fail the script filter (a "hex socket organizer" failed every iteration), and code-gen hears why a script failed, not about GPU or LLM errors;
    • the code-gen prompt states the one-solid rule and that a cut only removes existing material: live, that organizer brief went from sockets cut into an empty cavity (0 of 4 iterations) to right on the first try;
    • STL exports tessellate curves smoothly.
  • Blender: renders lost the hard black band across the top (the clip range and floor now scale with the object), stills are framed on the whole object (a robot was framed from the knees down), and exported GLBs no longer carry the startup file's Cube as a second scene.

  • chimera update-check told a ComfyUI ahead of the latest release (0.41.1 against the v0.41.0 release) to update.

  • Local models on Ollama 0.40 returned empty replies: thinking used up the whole token budget, so every CAD-loop iteration failed.

  • FLUX.2 jobs failed partway through a session on ComfyUI 0.40 (Windows error 1450). They now free ComfyUI's memory first.

  • After a brand or default model changed, replay could silently render on a different model.

  • auto_generate --pipeline mesh3d --model X loaded X as the 3D checkpoint, and a second loop run reused the first run's seeds and replaced its winner.

  • auto_generate.py exited 0 even when nothing passed.

Validated live on 2026-10-15 on ComfyUI 0.40 and Ollama 0.40 (RTX 5090). The run covered every new default, the alternates that stayed, the image, CAD and 3D self-correction loops on local models, replay, and doctor, and checked that each output was freshly created. All the README media were then regenerated on the new defaults, including the whole example-brand showcase, which now uses Qwen-Image-2.1. That pass is what surfaced most of the fixes above. After the GPU fix, re-checked on ComfyUI 0.41.1: doctor, image, product edit with SeedVR2, a sound effect, Wan 2.2 video, TRELLIS.2, the image loop on each GPU tier and the CAD loop on the large one. A second review's fixes were re-checked live too: the image loop with a repeated seed, the CAD loop, image replay, ACE-Step with lyrics (vocals confirmed by ear at 40 s), TRELLIS.2 with --format stl, and a Blender finish export. After a third review: the image and CAD loops on the large tier, the endpoint check, replays of genuine v0.5.1 sidecars, a Wan clip started from a brand render with foley on it (and that foley replayed without PyAV), and a Blender render. Before release: the 34-case judge benchmark re-run on the final verdict parser (Qwen 3.8 27B 33.0 of 34 correct, up from 31.0; Qwen3-VL 8B 30.0, up from 29.3; 4B 27.3), each tier's image loop after the GPU-estimate fix, the mesh3d loop (TRELLIS.2, PASS on the first try), a fresh clone installed in a new Python 3.12 environment (new-brand, lint, doctor, a render and a pixel-identical replay), and 204 raw judge replies read against the parser, then 24,426 rewrites of them. 1172 GPU-free tests · ruff clean · cross-platform CI. Full details are in the CHANGELOG; every change since the last release: v0.5.1...v0.6.0.

🤖 Generated with Claude Code