Model guide
CoCo identifies each model’s architecture from the file contents; nothing needs to be set by hand. When you choose a model of a different architecture, steps, CFG and sampler switch to that architecture’s recommended values (see Steps, CFG and samplers).
Supported architectures
Section titled “Supported architectures”| Architecture | Notes | Details |
|---|---|---|
| SDXL family | Illustrious, NoobAI (including V-Pred), Pony, Animagine and others | SDXL family |
| Anima | An anime-focused architecture | Anima |
| FLUX.2 klein (4B, 9B) | Fast; also supports edit by instruction | FLUX |
| FLUX.1 (dev, schnell, Kontext) | Kontext supports edit by instruction | FLUX |
| Z-Image | Other architectures | |
| Krea 2 (Turbo) | Turbo version only | Other architectures |
| Qwen Image 2.1 | Can render text; also supports edit by instruction | Other architectures |
| SD 1.5 | Loads, but hasn’t been fully tested | Other architectures |
Single-file and split models
Section titled “Single-file and split models”- Single file: SDXL and SD 1.5 models are one complete file that goes in
checkpoints. The VAE is included. - Split: for the other architectures, the main model, VAE and text encoder are separate files. The main model goes in
diffusion_models; the VAE and text encoder go invaeandtext_encoders.
The VAE and text encoder each architecture needs:
| Architecture | VAE | Text encoder |
|---|---|---|
| Anima | qwen_image_vae |
qwen_3_06b_base |
| Z-Image | ae |
qwen_3_4b |
| Krea 2 | qwen_image_vae |
qwen3vl_4b_fp8_scaled |
| Qwen Image 2.1 | qwen_image_2.1_vae_bf16 |
qwen3vl_8b_int8_convrot |
| FLUX.2 klein 4B | flux2-vae |
qwen_3_4b |
| FLUX.2 klein 9B | flux2-vae |
qwen_3_8b_fp8mixed |
| FLUX.1 | ae |
t5xxl_fp16 (or the fp8 version), plus clip_l |
All of these can be downloaded under “VAE and text encoders” in ⚙ Settings → Resource folders → Download models; files you already have are shown by name.
- When no VAE or encoder is chosen in the generate panel, CoCo finds matching files in the asset library automatically (see Asset library).
- A text encoder can be swapped for a different precision of the same model (bf16, fp8, int8), but not for a different model.
- Z-Image works only with
ae, and Qwen Image 2.1 only with its own VAE.
Features by architecture
Section titled “Features by architecture”| Feature | SDXL | Anima | Z-Image | Krea 2 | Qwen Image 2.1 | FLUX.2 klein | FLUX.1 |
|---|---|---|---|---|---|---|---|
| Generate, LoRA | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Whole image, inpaint, face fix, upscale, outpaint | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Reference images (ControlNet) | ✓ | ||||||
| Edit by instruction | ✓ | ✓ | Kontext | ||||
| LoRA training | ✓ | ✓ |
Model licences
Section titled “Model licences”CoCo doesn’t ship with generation models. Every model (including community fine-tunes) has its own licence; check it on the model’s page before use. The licences of some common models:
| Model | Licence |
|---|---|
| Anima | CircleStone Labs Non-Commercial License: the model itself may not be used commercially; generated images are unrestricted |
| Qwen Image 2.1 | Qwen Research License: non-commercial only |
| FLUX.2 klein 4B | Apache 2.0 |
| FLUX.2 klein 9B | Non-commercial licence |
| FLUX.1 dev, Kontext | Non-commercial licence |