Skip to content

v0.6.4 — the memory-guard release

Latest

Choose a tag to compare

@whartons whartons released this 13 Nov 03:17
· 1 commit to main since this release
439b4db

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.safetensors from Comfy-Org/flux2-dev (split_files/text_encoders/) into models/text_encoders/, or rename clip_name in the FLUX.2 templates if you keep the bf16 one.
  • The default relight runs on flux2-dev-nvfp4-mixed.safetensors (from black-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 --upscale needs ComfyUI 0.45 and the RelayVSR files from Comfy-Org/RelayVSR: relayvsr_x2_lightx2v_int8_convrot in diffusion_models/, relayvsr_flash_vae_x2_fp32 in vae/.
  • ComfyUI-LTXVideo moves to 220e1d4. git checkout 220e1d4, then pip install -r requirements.txt (it adds colour-science and OpenImageIO). Upstream fixed STG itself, so drop the local patch if you applied it.
  • freecad-mcp moves to v0.1.30 (ec8abf2). Re-copy addon/FreeCADMCP/ into FreeCAD's Mod/, 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 doctor shows the room, and CHIMERA_SKIP_MEMORY_CHECK=1 runs 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 --upscale is 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 doctor and chimera free report 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.