csprun/FireRed-Image-Edit-1.1-ComfyUI_mxfp8

This is an mxfp8 quantization of FireRedTeam/FireRed-Image-Edit-1.1-ComfyUI which can be used directly in ComfyUI. All credit goes to FireRedTeam (obviously), and the license here follows theirs.

Requirement: a Blackwell GPU (e.g. Nvidia GeForce RTX 5090 32GB) is required in order to actually take advantage of this quantization.

Quantization Details

Summary

• Source: FireRed-Image-Edit-1.1-transformer.safetensors — 40.86 GB (bf16)
• Output: FireRed-Image-Edit-1.1-transformer_mxfp8.safetensors — 21.11 GB (52% of source)
• Format: mxfp8 — block-32 microscaling FP8 (E4M3 + E8M0 block scales); ComfyUI-native (contains comfy_quant metadata, no custom nodes needed)
• Method: post-training round-to-nearest, weight-only

disposition	layers	parameters
Quantized (mxfp8)	840	20.38 B (100%)
Kept in bf16	1093	0.05 B (0%)
Total	1933	20.43 B

Quantized layers (mxfp8)

layer pattern	count	example shape
transformer_blocks.N.attn.add_k_proj.weight	60	(3072, 3072)
transformer_blocks.N.attn.add_q_proj.weight	60	(3072, 3072)
transformer_blocks.N.attn.add_v_proj.weight	60	(3072, 3072)
transformer_blocks.N.attn.to_add_out.weight	60	(3072, 3072)
transformer_blocks.N.attn.to_k.weight	60	(3072, 3072)
transformer_blocks.N.attn.to_out.N.weight	60	(3072, 3072)
transformer_blocks.N.attn.to_q.weight	60	(3072, 3072)
transformer_blocks.N.attn.to_v.weight	60	(3072, 3072)
transformer_blocks.N.img_mlp.net.N.proj.weight	60	(12288, 3072)
transformer_blocks.N.img_mlp.net.N.weight	60	(3072, 12288)
transformer_blocks.N.img_mod.N.weight	60	(18432, 3072)
transformer_blocks.N.txt_mlp.net.N.proj.weight	60	(12288, 3072)
transformer_blocks.N.txt_mlp.net.N.weight	60	(3072, 12288)
transformer_blocks.N.txt_mod.N.weight	60	(18432, 3072)

Kept in bf16 (not quantized)

layer pattern	count	example shape	reason/heuristic
img_in.bias	1	(3072,)	non-weight tensor (bias)
norm_out.linear.bias	1	(6144,)	non-weight tensor (bias)
proj_out.bias	1	(64,)	non-weight tensor (bias)
time_text_embed.timestep_embedder.linear_1.bias	1	(3072,)	non-weight tensor (bias)
time_text_embed.timestep_embedder.linear_2.bias	1	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.add_k_proj.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.add_q_proj.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.add_v_proj.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.to_add_out.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.to_k.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.to_out.N.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.to_q.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.to_v.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.img_mlp.net.N.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.img_mlp.net.N.proj.bias	60	(12288,)	non-weight tensor (bias)
transformer_blocks.N.img_mod.N.bias	60	(18432,)	non-weight tensor (bias)
transformer_blocks.N.txt_mlp.net.N.bias	60	(3072,)	non-weight tensor (bias)
transformer_blocks.N.txt_mlp.net.N.proj.bias	60	(12288,)	non-weight tensor (bias)
transformer_blocks.N.txt_mod.N.bias	60	(18432,)	non-weight tensor (bias)
txt_in.bias	1	(3072,)	non-weight tensor (bias)
transformer_blocks.N.attn.norm_added_k.weight	60	(128,)	1D weight (norm — not a Linear)
transformer_blocks.N.attn.norm_added_q.weight	60	(128,)	1D weight (norm — not a Linear)
transformer_blocks.N.attn.norm_k.weight	60	(128,)	1D weight (norm — not a Linear)
transformer_blocks.N.attn.norm_q.weight	60	(128,)	1D weight (norm — not a Linear)
txt_norm.weight	1	(3584,)	1D weight (norm — not a Linear)
img_in.weight	1	(3072, 64)	min dim < 1024 (small in/out projection)
proj_out.weight	1	(64, 3072)	min dim < 1024 (small in/out projection)
time_text_embed.timestep_embedder.linear_1.weight	1	(3072, 256)	min dim < 1024 (small in/out projection)
norm_out.linear.weight	1	(6144, 3072)	sensitive name: matched 'norm_out'
time_text_embed.timestep_embedder.linear_2.weight	1	(3072, 3072)	sensitive name: matched 'embed'
txt_in.weight	1	(3072, 3584)	sensitive name: matched 'txt_in'

Quality — reload verification

Round-trip dequantization error was computed over 64 sampled quantized layers (reconstructed exactly as ComfyUI's loader does). Relative L1 error: mean 0.0224, median 0.0224, max 0.0226 (~0.02 is near the mxfp8 representational floor).

Loading in ComfyUI

Load with Load Diffusion Model (or similar); the comfy_quant metadata is auto-detected and the weights stay quantized in VRAM (no custom nodes required). Requires a Blackwell GPU (mxfp8 tensor-core compute).