89 lines
4.5 KiB
Markdown
89 lines
4.5 KiB
Markdown
# Example workflows
|
|
|
|
One ready-to-run workflow per model family. Each file is the same three-node graph (Config Builder -> Sampler Grid -> Grid Dashboard) with the Builder already filled in for that model: model, text encoder, CLIP type, VAE, sampler/scheduler/steps/cfg axes and a prompt.
|
|
|
|
## How to load one
|
|
|
|
1. Download the model files listed for the example and put each in the folder shown (under `ComfyUI/models/`). The official Comfy-Org workflow templates and Hugging Face repos for each family ship exactly these files.
|
|
2. Drag the `.json` onto the ComfyUI canvas.
|
|
3. Open the Config Builder. If your file names differ, pick yours in the model / text encoder / VAE dropdowns. A file you do not have is shown as missing rather than silently replaced.
|
|
4. Queue the prompt. Results appear in the Dashboard node.
|
|
|
|
Times are approximate on an RTX 5060 Ti 16 GB, not counting the first model load.
|
|
|
|
## SDXL / Illustrious (`uscg-sdxl-illustrious.json`)
|
|
|
|
| File | Folder |
|
|
|---|---|
|
|
| `sd_xl_base_1.0.safetensors` | `models/checkpoints/` |
|
|
|
|
* **What the grid tests:** 2 samplers (dpmpp_2m, euler) x 2 schedulers (karras, normal) x cfg 5 and 7 x 2 prompts at 25 steps, 1024x1024. **16 images**
|
|
* **Time:** TBD
|
|
* **Notes:** the checkpoint has the VAE and CLIP built in, so no extra files. Swap in any Illustrious/SDXL checkpoint.
|
|
|
|
## Anima (`uscg-anima.json`)
|
|
|
|
| File | Folder |
|
|
|---|---|
|
|
| `anima-base-v1.0.safetensors` | `models/diffusion_models/` |
|
|
| `qwen_3_06b_base.safetensors` | `models/text_encoders/` |
|
|
| `qwen_image_vae.safetensors` | `models/vae/` |
|
|
|
|
* **What the grid tests:** 2 samplers (er_sde, euler) x simple scheduler x cfg 4 and 5 at 30 steps, 1024x1024. **4 images**
|
|
* **Time:** TBD
|
|
* **Notes:** the CLIP type stays `stable_diffusion`; there is no separate Anima option and ComfyUI picks the Anima text encoder from the Qwen3-0.6B file.
|
|
|
|
## Z-Image Turbo (`uscg-z-image-turbo.json`)
|
|
|
|
| File | Folder |
|
|
|---|---|
|
|
| `z_image_turbo_bf16.safetensors` | `models/diffusion_models/` |
|
|
| `qwen_3_4b.safetensors` | `models/text_encoders/` |
|
|
| `ae.safetensors` | `models/vae/` |
|
|
|
|
* **What the grid tests:** res_multistep / simple at 8 and 10 steps, cfg 1, 1024x1024. **2 images**
|
|
* **Time:** TBD
|
|
* **Notes:** CLIP type `lumina2`. Model Sampling is set to AuraFlow with shift 3, as in the official template.
|
|
|
|
## Krea 2 Turbo (`uscg-krea2-turbo.json`)
|
|
|
|
| File | Folder |
|
|
|---|---|
|
|
| `krea2_turbo_fp8_scaled.safetensors` | `models/diffusion_models/` |
|
|
| `qwen3vl_4b_fp8_scaled.safetensors` | `models/text_encoders/` |
|
|
| `qwen_image_vae.safetensors` | `models/vae/` |
|
|
|
|
* **What the grid tests:** euler / simple at 6 and 8 steps, cfg 1, 1024x1024. **2 images**
|
|
* **Time:** TBD
|
|
* **Notes:** CLIP type `krea2`.
|
|
|
|
## Flux.2 Klein 9B (`uscg-flux2-klein-9b.json`)
|
|
|
|
| File | Folder |
|
|
|---|---|
|
|
| `flux-2-klein-9b-fp8.safetensors` | `models/diffusion_models/` |
|
|
| `qwen_3_8b_fp8mixed.safetensors` | `models/text_encoders/` |
|
|
| `flux2-vae.safetensors` | `models/vae/` |
|
|
|
|
* **What the grid tests:** euler at 4 and 6 steps, 1024x1024, each with the Advanced Sampling pipeline (CFG guider + Flux2 scheduler) and without it (plain KSampler, simple scheduler): the Builder always grids Advanced Sampling on vs off. **4 images**
|
|
* **Time:** TBD
|
|
* **Notes:** Klein is distilled, so cfg is 1; the official template uses the CFG guider plus the Flux2 scheduler, which is the Advanced Sampling half of the grid. CLIP type `flux2`.
|
|
|
|
## MOSS-Image (`uscg-moss-image.json`)
|
|
|
|
| File | Folder |
|
|
|---|---|
|
|
| `moss_image_sft_int8_convrot.safetensors` | `models/diffusion_models/` |
|
|
| `qwen3.5_9b_w4a8_gptq.safetensors` | `models/text_encoders/` |
|
|
| `flux2-vae.safetensors` | `models/vae/` |
|
|
|
|
* **What the grid tests:** two config arrays with the same prompt and settings (er_sde / simple, 25 steps, cfg 2.5): one with Skip Layer Guidance on, one off. **2 images**
|
|
* **Time:** TBD
|
|
* **Notes:** SLG uses the official recipe, single block 22 for the first 0-50 % of the steps. Compare the two images to see what it does. CLIP type `moss_image`. A smaller `flux2-vae_small.safetensors` VAE also exists and can be picked in the Builder instead. The 9B text encoder is large; on a 16 GB card expect model swapping.
|
|
|
|
## Make your own
|
|
|
|
* Edit the Builder: add LoRAs or a whole LoRA folder, more samplers, schedulers, step counts, cfg values, resolutions or prompts. Every list is multiplied together.
|
|
* Raise **Additional Random Seeds** in Run Settings to render extra seeds per config, handy for telling what the model does from what a LoRA does.
|
|
* Add another config array to compare models or settings side by side (the MOSS example does this for SLG).
|