ComfyUI Model Tools

Utility scripts for packaging models for ComfyUI.

Merge Safetensors

merge_safetensors.py: Merge all safetensors files in a directory into a single safetensors file. If duplicate keys exist, they are skipped with a warning:

python.exe merge_safetensors.py ".\source\folder\path" "target_file_path.safetensors"

Convert SeedVR2 to Safetensors

seedvr2_convert.py: Convert a ByteDance SeedVR2 .pth checkpoint into ComfyUI-loadable safetensors in fp16, fp8_e4m3fn, mxfp8, nvfp4, or int8. Each --job writes one precision; an optional PRECISION:OUT:SHA256 verifies the written file:

python seedvr2_convert.py --src seedvr2_ema_3b.pth --cond pos_emb.pt,neg_emb.pt \
    --job fp16:seedvr2_3b_fp16.safetensors \
    --job nvfp4:seedvr2_3b_nvfp4.safetensors

The mxfp8, nvfp4, and int8 modes require the comfy-kitchen package; the script exits with a message naming it if it is not installed.

Reproduce Qwen release artifacts

convert_qwen3vl.py and convert_qwen35.py are Rev0 deterministic release recipes. They accept only embedded, SHA-pinned original Qwen checkpoints and always verify the complete output file against the corresponding artifact published by Comfy-Org on Hugging Face. Unsupported models or source revisions fail instead of inheriting a recipe from another family member.

Qwen3-VL supports the released 4B BF16/FP8-scaled artifacts and the released 8B BF16/FP8-scaled/NVFP4 artifacts:

python convert_qwen3vl.py --src Qwen3-VL-4B-Instruct-snapshot \
    --job bf16:qwen3vl_4b_bf16.safetensors \
    --job fp8_scaled:qwen3vl_4b_fp8_scaled.safetensors

Qwen3.5 selects exactly one model-specific recipe from the pinned source shard hashes: 2B is copied byte-for-byte, 4B is merged without metadata, and 9B is merged with {"format": "pt"} metadata.

python convert_qwen35.py --src Qwen3.5-4B-snapshot \
    --out qwen3.5_4b_bf16.safetensors

See QWEN_CONVERTER_PROVENANCE.md for the pinned source revisions, failed pre-Rev0 reconstruction findings, and all eight canonical output hashes proved by full regeneration.

Quantize to int8-convrot

Dry-run first (good idea to do this on a new architecture) — prints the plan, writes nothing

python quant_int8_auto.py model_bf16.safetensors --dry-run

Quantize (defaults: absmax, min-gemm 256)

python quant_int8_auto.py model_bf16.safetensors model_int8_convrot.safetensors

args

--dry-run — plan only (grouped quantize list + what's left behind, with reasons)
--exclude RE / --include RE — regex overrides on layer base names
--min-gemm N — skip layers with min(N,K) < N (default 256)
--mseclip — MSE-optimal clip instead of absmax, usually lower weight error, experimental
--downcast-fp32 — shrink stray fp32 passthrough to the compute dtype
--warn-thresh F — warn on any quantized layer over F% relerr (default 2.0)
--verify-report PATH — dump the full per-layer (relerr, cos, gs) table