nanaj's picture
Upload folder using huggingface_hub
5e296ef verified
|
Raw
History Blame Contribute Delete
4.16 kB

Model Quantizer Presets

Preset review date: 2026-07-11. The GUI favors a proven model recipe over applying one quantization format to every architecture.

Quality Presets

Preset Intended use Runtime
INT8 ConvRot Balanced Default. Simple row-wise INT8 with fixed ConvRot group 256 Native current ComfyUI, Turing+
INT8 ConvRot Learned Higher reconstruction quality with Prodigy and 2,000 iterations Native current ComfyUI, Turing+
INT8 Tensorwise Models whose dimensions or public recipes do not support fixed ConvRot well Native current ComfyUI, Turing+
FP8 Simple Fast, low-memory FP8 conversion Native current ComfyUI; best acceleration on Ada+
FP8 Learned Balanced or high-quality learned FP8 Native current ComfyUI; best acceleration on Ada+
FP8 Compatibility Full-precision matmul metadata for known quality/compatibility exceptions Native current ComfyUI fallback path
INT8 Blockwise Experimental block layout only QuantOps required
MXFP8 / NVFP4 Expert Blackwell formats Native current ComfyUI on Blackwell

full_precision_matrix_mult is false for native compute presets. It is true only for the compatibility preset and the Qwen Image quality-first recipe. The removed aggressive Z-Image NVFP4 preset is mapped to FP8 Compatibility when an old config is loaded.

Model Defaults

Model preset Default Important protected behavior
FLUX.2 INT8 ConvRot Balanced Modulation, guidance, time, final, and input layers
FLUX 2 Klein INT8 ConvRot Balanced Flux-sensitive layers except the absent guidance input
Krea 2 Raw/Turbo INT8 ConvRot Balanced First/last, time, input, final, and text-fusion layers; no layer-config ConvRot bypass
Ideogram 4 INT8 ConvRot Balanced Image indicator, time, condition, AdaLN, final, and input projections
WAN 2.1/2.2 INT8 ConvRot Balanced Embeddings, condition/audio paths, packers, heads, and adapters
LTX 2/2.3 INT8 ConvRot Balanced Boundary blocks, connectors, VAE, vocoder, and gates
Gemma 4 INT8 ConvRot Balanced Embeddings, K/V projections, audio/vision, and multimodal projectors
Qwen3.5 INT8 ConvRot Balanced Layers 0 and 23/31/63, embeddings, MTP, and complete visual stack
T5-XXL INT8 ConvRot Balanced Norm/bias exclusions, decoder removal, and input scales
Radiance / distilled / NeRF INT8 ConvRot Balanced Their upstream high-precision boundary and guidance layers
Boogu-Image INT8 Tensorwise Embeddings plus norm1.linear and norm_out; avoids incompatible block/ConvRot widths
Qwen Image / Edit FP8 Learned Reference Learned rounding and full-precision matmul; Qwen-sensitive norms and boundary layers
Z-Image / Refiner FP8 Compatibility Z-Image norms, embeddings, final layers, and refiner blocks
Anima Base/Turbo FP8 Compatibility Initial blocks, AdaLN, adapters, time/input/final layers
FLUX.1 FP8 Learned Reference Flux-sensitive layers and input-scale tensors
ERNIE Image FP8 Learned Reference Author-recommended time, AdaLN, first attention/MLP, final, and text projection layers
Hunyuan Video / LENS FP8 Learned Balanced Existing upstream sensitive-layer filters; no public ConvRot validation yet
Generic Text Encoder FP8 Learned Balanced Input-scale behavior only; no architecture-specific exclusion claim

Safety Rules

  • Fixed ConvRot group 256 is the stable default. Dynamic groups remain an explicit experiment.
  • The GUI blocks detected quantized inputs from the Quantize workflow unless the expert override is enabled.
  • W4A4 ConvRot is not listed yet: ComfyUI can load the new layout, but convert_to_quant cannot currently create it.
  • Choose the model preset before changing the quality preset so high-precision exclusions remain active.

Upstream references: convert_to_quant, model filter PR 27, ComfyUI, and comfy-kitchen.