| File |
Goes in |
File name |
| Main model |
diffusion_models |
a Z-Image main model |
| VAE |
vae |
ae (the only one that works) |
| Text encoder |
text_encoders |
qwen_3_4b (the fp8 version works too) |
| Item |
Default |
Notes |
| Steps |
28 |
Usually 28–50 |
| CFG |
4 |
Usually 3–5 |
| Sampler |
Euler |
|
| shift |
6.0 |
The official value for the Base version; 3.0 for Turbo |
- With Z-Image’s distillation LoRA, 8 steps at CFG 1 produce a complete image (the style changes).
- In Whole image refines, Z-Image changes more than other models; lower the amount of change to keep more of the original.
Krea 2 comes in two versions, Turbo and Raw; CoCo supports Turbo only. Raw is the undistilled base model, intended officially for LoRA training rather than for generating finished images.
| File |
Goes in |
File name |
| Main model |
diffusion_models |
The fp8 version (krea2_turbo_fp8_scaled) or NVFP4 version (krea2_turbo_nvfp4) of Krea 2 Turbo |
| VAE |
vae |
qwen_image_vae |
| Text encoder |
text_encoders |
qwen3vl_4b_fp8_scaled |
| Item |
Default |
| Steps |
8 |
| CFG |
1 |
| Sampler |
Euler |
| shift |
3.16 |
- Both
qwen_image_vae and wan_2.1_vae match Krea 2; CoCo prefers qwen_image_vae.
- The fp8 and NVFP4 versions both work. With the same seed they give different compositions; the fp8 version uses less graphics memory, the NVFP4 version has a smaller file.
- The int8 and mxfp8 versions can’t be used yet.
| File |
Goes in |
File name |
| Main model |
diffusion_models |
a Qwen Image 2.1 main model (for example the int8 version) |
| VAE |
vae |
qwen_image_2.1_vae_bf16 (specific to 2.1; no other VAE works) |
| Text encoder |
text_encoders |
qwen3vl_8b_int8_convrot |
| Item |
Default |
Notes |
| Steps |
40 |
|
| CFG |
1 |
|
| Sampler |
Euler |
|
| shift |
2.0 |
The value at 1024×1024; other resolutions adjust automatically |
- Can render text in the image.
- Width and height must be multiples of 32; other sizes are drawn slightly larger and scaled back.
- Also used for Edit by instruction in Refine and for making slider pairs.
- In Whole image refines, you may need a higher amount of change than with other models before prompt changes show.
- Licensed under the Qwen Research License: non-commercial only.
SD 1.5 models load and generate, but haven’t been fully tested. ControlNet models for SD 1.5 can’t be used.
The main models of Z-Image, Krea 2 and Qwen Image 2.1 are large. When they don’t fit on the card, CoCo streams them automatically instead of failing, but enough system memory is needed (see Troubleshooting). “Store the main model in fp8” makes bf16 main models take about half the memory.