Feature Idea
Is your feature request related to a problem? Please describe.
While the native INT8 and INT8-ConvRot support introduced in ComfyUI is a massive quality and performance milestone, it currently throws errors or is unsupported on several newer/modern model architectures.
For example, loading LTX-2.3 in native INT8-ConvRot fails with errors like:
ValueError: Unknown quantization format for layer transformer_blocks.0.attn1.to_q
For users on Ampere architecture GPUs (such as the RTX 3080 Ti or 3090), FP8 performance is sub-optimal compared to newer Ada Lovelace or Blackwell architectures. However, INT8-ConvRot provides an exceptional quality-to-performance ratio on these GPUs, delivering fast generation speeds while maintaining high visual quality without the heavy degradation often seen in standard FP8 or row-wise INT8.
Currently, users have to rely on third-party custom nodes to get these models running in INT8-ConvRot, which fragments the workflow.
Describe the solution you'd like
We would love to see native INT8-ConvRot support expanded and optimized in ComfyUI's native loaders to fully support the following modern model architectures:
Flux-2-Klein Variants:
flux-2-klein-9b
flux-2-klein-9b-kv
Qwen Image Series:
Qwen-Image-2512
Qwen-Image-Edit-2511
LTX Video Series:
LTX-2.3-22b
SeedVR2 Series:
seedvr2_3b
seedvr2_ema_7b
seedvr2_ema_7b_sharp
Ensuring the native compiler/loader can parse the quantization format for these attention layers, linear projections, and biases would make ComfyUI the ultimate tool for running these heavy models on 30-series cards.
Describe alternatives you've considered
Using custom nodes to load these checkpoints. While this works, native implementation in ComfyUI is highly preferred for stability, official updates, and seamless integration with other native features (like native LoRA loading without re-quantization bugs).
Additional context
Tests within the community consistently show that INT8-ConvRot outperforms traditional FP8 in both speed (especially on RTX 30-series) and structural/textural quality preservation. Native support for these specific architectures would democratize the usage of these high-end models for a huge portion of the user base running mid-to-high-end 30-series GPUs.
Existing Solutions
No response
Other
No response
Feature Idea
Is your feature request related to a problem? Please describe.
While the native INT8 and
INT8-ConvRotsupport introduced in ComfyUI is a massive quality and performance milestone, it currently throws errors or is unsupported on several newer/modern model architectures.For example, loading LTX-2.3 in native INT8-ConvRot fails with errors like:
ValueError:
Unknown quantization format for layer transformer_blocks.0.attn1.to_qFor users on Ampere architecture GPUs (such as the RTX 3080 Ti or 3090), FP8 performance is sub-optimal compared to newer Ada Lovelace or Blackwell architectures. However, INT8-ConvRot provides an exceptional quality-to-performance ratio on these GPUs, delivering fast generation speeds while maintaining high visual quality without the heavy degradation often seen in standard FP8 or row-wise INT8.
Currently, users have to rely on third-party custom nodes to get these models running in INT8-ConvRot, which fragments the workflow.
Describe the solution you'd like
We would love to see native INT8-ConvRot support expanded and optimized in ComfyUI's native loaders to fully support the following modern model architectures:
Flux-2-Klein Variants:
flux-2-klein-9bflux-2-klein-9b-kvQwen Image Series:
Qwen-Image-2512Qwen-Image-Edit-2511LTX Video Series:
LTX-2.3-22bSeedVR2 Series:
seedvr2_3bseedvr2_ema_7bseedvr2_ema_7b_sharpEnsuring the native compiler/loader can parse the quantization format for these attention layers, linear projections, and biases would make ComfyUI the ultimate tool for running these heavy models on 30-series cards.
Describe alternatives you've considered
Using custom nodes to load these checkpoints. While this works, native implementation in ComfyUI is highly preferred for stability, official updates, and seamless integration with other native features (like native LoRA loading without re-quantization bugs).
Additional context
Tests within the community consistently show that INT8-ConvRot outperforms traditional FP8 in both speed (especially on RTX 30-series) and structural/textural quality preservation. Native support for these specific architectures would democratize the usage of these high-end models for a huge portion of the user base running mid-to-high-end 30-series GPUs.
Existing Solutions
No response
Other
No response