This HuggingFace package repackages a Qwen3-VL-32B “ultra uncensored heretic” fork into ComfyUI-compatible H3 conditioning encoder checkpoints, offering both a full BF16 source and an INT8 ConvRot quantized build plus an optional generation tail. The practical aim is to let ComfyUI users run H3-style conditioning (language layers 0–49 + vision tower) on machines where the full 48 GiB BF16 encoder would otherwise be infeasible.
What Sets It Apart
- BF16 source + INT8 ConvRot alternative: The BF16 conditioning encoder is ~47.97 GiB (complete vision tower and language layers 0–49) while the INT8 ConvRot build reduces that to ~24.55 GiB by converting learned language matrices to learned row-wise INT8 ConvRot with group size 256, retaining 551 tensors in BF16 (notably the vision tower and norms). This gives a clear memory/performance trade-off for ComfyUI users.
- H3-tail workflow support: An optional ~7.09 GiB generation tail (layers 50–63, final norm, LM head) can be loaded transiently by H3 prompt enhancer nodes so the conditioning CLIP remains unchanged after prompt enhancement, enabling generation without duplicating the vision/token embedding state.
- Provenance and validation: The package is built from a pinned upstream revision and validated byte-for-byte for retained BF16 tensors; conversion used AdamW AdaRound ConvRot optimization rather than naive rounding to preserve accuracy in quantized matrices.
- Includes uncensoring edits from the upstream “Heretic” fork (ARA edits in layers 31–40), which the source reports reduced refusal rates and changed behavioral characteristics — a deliberate design choice that affects safety and outputs.
Who It's For and Trade-offs
Great fit if you run ComfyUI H3 workflows and need a working balance between fidelity and VRAM: use BF16 when you have ~48 GiB available for highest fidelity; use INT8 ConvRot on GPUs with ~24–32 GiB VRAM to enable H3 conditioning. The package was runtime-tested (example: RTX 5090, ~24.7 GiB allocated) but may rely on up-to-date ComfyUI/comfy-kitchen and recommended PyTorch/CUDA builds for optimal kernels. Look elsewhere if you require upstream model safety defaults (this build includes uncensoring edits that reduce refusal behavior) or need a turnkey Transformers generation repo—these are ComfyUI checkpoints, not a full generation framework. Also consider that quantization/ConvRot introduces subtle accuracy trade-offs despite careful AdaRound optimization.
Where It Fits
Use this when you want to experiment with H3-style conditioning, prompt-enhancement tails, or run Qwen3-VL-32B-based multimodal encoders inside ComfyUI without the full BF16 memory cost. Prefer the BF16 variant for fidelity-sensitive pipelines and the INT8 ConvRot variant for memory-limited interactive setups.
