8-step quality without the bloat.
Z-Image is an S3-DiT with a Qwen3-4B caption encoder. The distilled Z-Image-Turbo checkpoint generates in 8 steps with CFG off and lands between Anima and Krea 2 in both speed and detail. The non-distilled Z-Image base checkpoint uses the same engine and the same VAE/text encoder, but wants more steps.
- Text-to-image only (no editing)
- Turbo: built-in defaults 8 steps, guidance 1.0 (CFG off)
- Base: pass
--steps/--guidance-scaleexplicitly (see below) - int8-convrot variant available for both
Commands below assume a dev checkout (
./thenoise.sh,.venv/bin/python). On a portable bundle use./bin/thenoiseand./bin/python3instead.
| Spec | Data |
|---|---|
| Architecture | S3-DiT (flow matching) |
| Download size | ~21 GB (bf16 DiT + Flux VAE + Qwen3-4B) |
| VAE | Flux VAE |
| Text encoder | Qwen3-4B (caption encoder) |
| Editing | — |
| KV cache | — |
| Default settings | 1024×1024, 8 steps, guidance 1.0, euler sampler |
| Checkpoint | Repo | Steps |
|---|---|---|
--variant turbo (default) |
Comfy-Org/z_image_turbo |
8 (built-in default) |
--variant base |
Comfy-Org/z_image |
your choice (not distilled) |
For the base checkpoint the engine's built-in defaults target Turbo, so pass
explicit --steps and --guidance-scale on every run; the upstream
Z-Image page documents the
reference settings.
Ready-made Lemonade recipes for this model - each pins the checkpoints and the generation defaults:
- Z-Image-Turbo.json - Turbo BF16, 8 steps
- Z-Image-Turbo-INT8.json - Turbo INT8-ConvRot, 8 steps
- Z-Image.json - Base BF16, 28 steps, CFG 3
- Z-Image-INT8.json - Base INT8-ConvRot, 28 steps, CFG 3
- Turbo @ 8 steps: ~10 s (1024×768)
Generated - Turbo:
# Turbo (default)
.venv/bin/python scripts/download.py --model zimage
# non-distilled base
.venv/bin/python scripts/download.py --model zimage --variant base
# int8-convrot DiT instead of bf16 (either variant)
.venv/bin/python scripts/download.py --model zimage --int8-convrotEach fetch lands in ./models/zimage/ with the DiT under
split_files/diffusion_models/; the Flux VAE and Qwen3-4B encoder are shared
between both variants.
Generate (Turbo):
./thenoise.sh generate \
--dit ./models/zimage/split_files/diffusion_models/z_image_turbo_bf16.safetensors \
--vae ./models/zimage/split_files/vae/ae.safetensors \
--text-encoder ./models/zimage/split_files/text_encoders/qwen_3_4b.safetensors \
--prompt "a snowy mountain village at night" \
--out village.pngGenerate (base — explicit schedule):
./thenoise.sh generate \
--dit ./models/zimage/split_files/diffusion_models/z_image_bf16.safetensors \
--vae ./models/zimage/split_files/vae/ae.safetensors \
--text-encoder ./models/zimage/split_files/text_encoders/qwen_3_4b.safetensors \
--prompt "a snowy mountain village at night" \
--steps 28 --guidance-scale 3.0 \
--out village_base.pngServe over HTTP with the web UI (open http://localhost:8000/):
./thenoise.sh serve \
--dit ./models/zimage/split_files/diffusion_models/z_image_turbo_bf16.safetensors \
--vae ./models/zimage/split_files/vae/ae.safetensors \
--text-encoder ./models/zimage/split_files/text_encoders/qwen_3_4b.safetensors \
--host 127.0.0.1 --port 8000For every flag (size, steps, seed, LoRAs, upscaling, post-processing), see the CLI reference and the HTTP API reference.