Z-Image-Turbo ยท FP8 (transformer)

FP8 (float8_e4m3fn) quantization of the Z-Image DiT (distilled turbo) transformer from Tongyi-MAI/Z-Image-Turbo. Quantized by axteam.

File Size
Z-Image-Turbo-transformer-fp8_e4m3fn.safetensors 6.2 GB

What this is

  • Only the transformer (DiT) is quantized to float8_e4m3fn (2-D weight tensors; norms and biases stored in BF16, converted from the original FP32). About 25% of the size of the original FP32 transformer (24.6 GB), or half the size of a BF16 copy.
  • VAE and text encoder(s) are NOT included โ€” use the ones from the original Tongyi-MAI/Z-Image-Turbo.

Checked on GPU (diffusers)

On 2026-11-02 the FP8 values of this file (the same values ComfyUI loads; upcast to BF16 for this run) were loaded into the transformer of the original Tongyi-MAI/Z-Image-Turbo pipeline (diffusers 0.42.1, NVIDIA B200; all tensor names matched the original transformer โ€” 0 missing, 0 unexpected) and compared with the original pipeline run in BF16 at the same prompt and seed (1024ร—1024, 9 steps, guidance 0.0).

Prompt Pixel PSNR vs original (BF16 run) Mean abs. pixel diff
A red apple on a wooden table, studio lighting, photorealistic 37.4 dB 2.4 / 255
A cozy cafe street in Seoul at night with warm lights, illustration 19.2 dB 13.7 / 255

Both images match the BF16 ones closely, with the model's recommended Turbo settings: the apple is practically identical and the street scene keeps the same composition, with small differences in details such as the sign lettering; clean, normal quality. Pixel PSNR is strict for diffusion models: a small weight change shifts the sampling path, so a detailed scene can change its layout while the subject stays correct.

Requirements

Runtime ComfyUI (Load Diffusion Model node, float8_e4m3fn single file). Not run in ComfyUI here (no ComfyUI on this server); the FP8 weights were checked with diffusers โ€” see Checked on GPU above.
Also needed the Qwen3 text encoder (8.0 GB) and VAE from the base repo.
Memory this file is 6.2 GB; add the text encoder(s) and VAE above; the base model card says it fits in 16 GB VRAM.

Usage (ComfyUI)

  1. Download the FP8 file into ComfyUI's diffusion-model folder:
hf download axteam/Z-Image-Turbo-FP8 Z-Image-Turbo-transformer-fp8_e4m3fn.safetensors --local-dir ComfyUI/models/diffusion_models
  1. Get the VAE and text encoder(s) from the original Tongyi-MAI/Z-Image-Turbo repo (or the matching ComfyUI repackaged files) and put them in ComfyUI/models/vae/ and ComfyUI/models/text_encoders/.
  2. In your workflow, load this file with the Load Diffusion Model node (UNETLoader) โ€” leave weight_dtype at default (the file is already float8_e4m3fn) โ€” and wire the original VAE / text-encoder loaders as in the base model's reference workflow.
  3. Sampler settings (steps, CFG/guidance, resolution) are the same as for the original model.

The FP8 file is a plain cast of the transformer weights (no scaling), the common ComfyUI fp8_e4m3fn convention; norms and biases are stored in BF16 (the original is FP32).

License

Quantized derivative of Tongyi-MAI/Z-Image-Turbo. Original weights belong to the upstream authors; the base model's license applies to this derivative.


ํ•œ๊ตญ์–ด ์š”์•ฝ

Tongyi-MAI/Z-Image-Turbo์˜ transformer(DiT) FP8 ์–‘์žํ™” ๋นŒ๋“œ (ComfyUI์šฉ ๋‹จ์ผ ํŒŒ์ผ) ยท ์ œ์ž‘ axteam

๊ฐœ์š”

ํ•ญ๋ชฉ ๋‚ด์šฉ
์›๋ณธ ๋ชจ๋ธ Tongyi-MAI/Z-Image-Turbo
ํ˜•์‹ float8_e4m3fn ๋‹จ์ผ safetensors (ComfyUI์—์„œ ํ”ํžˆ ์“ฐ๋Š” FP8 ํ˜•์‹)
ํŒŒ์ผ Z-Image-Turbo-transformer-fp8_e4m3fn.safetensors 6.2 GB
ํฌํ•จ transformer(DiT)

์–‘์žํ™” ์ƒ์„ธ

  • ์–‘์žํ™” ์ ˆ์ฐจ: transformer์˜ 2D ๊ฐ€์ค‘์น˜๋ฅผ ์Šค์ผ€์ผ ์—†์ด float8_e4m3fn์œผ๋กœ ๋‹จ์ˆœ ํ˜•๋ณ€ํ™˜(cast) โ€” ComfyUI์—์„œ ํ”ํžˆ ์“ฐ๋Š” ๋ฐฉ์‹
  • BF16 ์ €์žฅ(์›๋ณธ FP32์—์„œ ๋ณ€ํ™˜): ์ •๊ทœํ™”์ธต, bias
  • ํฌ๊ธฐ: ์›๋ณธ FP32 transformer(24.6 GB)์˜ ์•ฝ 25% (BF16 ์‚ฌ๋ณธ์˜ ์ ˆ๋ฐ˜)
  • ๋ฏธํฌํ•จ: VAEยทํ…์ŠคํŠธ ์ธ์ฝ”๋” โ€” ์›๋ณธ ๋ฆฌํฌ์—์„œ ๋ฐ›์•„ ์‚ฌ์šฉ
  • ํ’ˆ์งˆ(์‹ค์ธก, 2026-11-02): diffusers 0.42.1๋กœ ์›๋ณธ ํŒŒ์ดํ”„๋ผ์ธ์„ BF16์œผ๋กœ ์‹คํ–‰ํ•œ ๊ทธ๋ฆผ๊ณผ, ๊ฐ™์€ ํŒŒ์ดํ”„๋ผ์ธ์˜ transformer์— ์ด ํŒŒ์ผ์˜ FP8 ๊ฐ’(ComfyUI๊ฐ€ ์ฝ๋Š” ๊ฒƒ๊ณผ ๊ฐ™์€ ๊ฐ’, ์ด ์ ๊ฒ€์—์„œ๋Š” BF16์œผ๋กœ ์˜ฌ๋ ค ๊ณ„์‚ฐ)์„ ๋„ฃ์–ด ๊ทธ๋ฆฐ ๊ทธ๋ฆผ์„ ๊ฐ™์€ ํ”„๋กฌํ”„ํŠธยท์‹œ๋“œ 2๊ฐœ๋กœ ๋น„๊ต(1024ร—1024) โ€” Turbo ๊ถŒ์žฅ ์„ค์ •์—์„œ ๋‘ ๊ทธ๋ฆผ ๋ชจ๋‘ ์›๋ณธ๊ณผ ๊ฐ€๊นŒ์›€, ์‚ฌ๊ณผ๋Š” ์‚ฌ์‹ค์ƒ ๊ฐ™๊ณ  ๊ฑฐ๋ฆฌ ์žฅ๋ฉด๋„ ๊ฐ™์€ ๊ตฌ๋„(๊ฐ„ํŒ ๊ธ€์ž ๋“ฑ ์„ธ๋ถ€๋งŒ ๋‹ค๋ฆ„), ์ •์ƒ ํ™”์งˆ, ํ”ฝ์…€ PSNR 37.4ยท19.2 dB (๋””ํ“จ์ „์€ ๊ฐ€์ค‘์น˜๊ฐ€ ์กฐ๊ธˆ๋งŒ ๋‹ฌ๋ผ๋„ ์ƒ์„ฑ ๊ฒฝ๋กœ๊ฐ€ ๋ฐ”๋€Œ์–ด ๋ณต์žกํ•œ ์žฅ๋ฉด์€ ๊ตฌ๋„๊ฐ€ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์Œ, ์œ„ ์˜์–ด ๋ณธ๋ฌธ ํ‘œ ์ฐธ๊ณ )

์š”๊ตฌ์‚ฌํ•ญ

  • ์‹คํ–‰ ํ™˜๊ฒฝ: ComfyUI โ€” Load Diffusion Model ๋…ธ๋“œ๋กœ float8_e4m3fn ๋‹จ์ผ ํŒŒ์ผ ๋กœ๋“œ (ComfyUI๋กœ๋Š” ์‹คํ–‰ ๋ฏธํ™•์ธ โ€” FP8 ๊ฐ€์ค‘์น˜๋Š” diffusers๋กœ ์ƒ์„ฑ ํ™•์ธ, ํ’ˆ์งˆ ํ•ญ๋ชฉ ์ฐธ๊ณ )
  • ์ถ”๊ฐ€๋กœ ํ•„์š”ํ•œ ํŒŒ์ผ: ์›๋ณธ ๋ฆฌํฌ์˜ Qwen3 ํ…์ŠคํŠธ ์ธ์ฝ”๋”(8.0 GB)์™€ VAE
  • ๋ฉ”๋ชจ๋ฆฌ: ์ด ํŒŒ์ผ 6.2 GB + ์œ„ ํ…์ŠคํŠธ ์ธ์ฝ”๋”ยทVAE โ€” ์›๋ณธ ์นด๋“œ ๊ธฐ์ค€ 16 GB VRAM์—์„œ ๋™์ž‘

์‚ฌ์šฉ๋ฒ• (ComfyUI)

  • ํŒŒ์ผ ๋ฐ›๊ธฐ: hf download axteam/Z-Image-Turbo-FP8 Z-Image-Turbo-transformer-fp8_e4m3fn.safetensors --local-dir ComfyUI/models/diffusion_models
  • ํ…์ŠคํŠธ ์ธ์ฝ”๋”ยทVAE: ์›๋ณธ ๋ฆฌํฌ์—์„œ ๋ฐ›์•„ ComfyUI/models/text_encoders/ยทComfyUI/models/vae/์— ๋ฐฐ์น˜
  • ๋ถˆ๋Ÿฌ์˜ค๊ธฐ: Load Diffusion Model ๋…ธ๋“œ(UNETLoader), weight_dtype์€ default (ํŒŒ์ผ์ด ์ด๋ฏธ float8_e4m3fn)
  • ์ƒ˜ํ”Œ๋Ÿฌ ์„ค์ •: ์Šคํ…ยทCFG(guidance)ยทํ•ด์ƒ๋„๋Š” ์›๋ณธ ๋ชจ๋ธ๊ณผ ๋™์ผ

๋ผ์ด์„ ์Šค

  • ๋ผ์ด์„ ์Šค: Apache-2.0 โ€” ์›๋ณธ ๋ชจ๋ธ์˜ ๋ผ์ด์„ ์Šค๋ฅผ ๊ทธ๋Œ€๋กœ ๋”ฐ๋ฆ„
  • ๊ถŒ๋ฆฌ: ์›๋ณธ ๊ฐ€์ค‘์น˜๋Š” Tongyi-MAI(Alibaba) ์†Œ์œ , ์ด ๋ฆฌํฌ๋Š” ์–‘์žํ™” ํŒŒ์ƒ๋ฌผ
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for axteam/Z-Image-Turbo-FP8

Quantized
(109)
this model

Collection including axteam/Z-Image-Turbo-FP8