trending/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-10-21 06:19:58 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
master
stable-diffusion.cpp/examples/server/recipes/defaults.md
T

125 lines
18 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Built-in generation defaults
Reviewed against model-publisher cards and reference implementations on 2026-10-20.
These are starting values, not limits on supported resolutions or generation settings.
Built-in step defaults above 20 are reduced to 20; the cited sources may recommend more steps.
Users can still choose a higher step count.
When several aspect ratios or ranges are published, the selected example or square bucket is noted below.
A dash means no recipe override: the available primary sources did not establish a fixed value, or the size depends on the input image.
Ordinary DiT recipes explicitly use Euler; PiD retains its LCM sampling path.
Guidance-distilled models use `cfg_scale: 1` to disable external CFG in sd.cpp. FLUX embedded guidance is stored separately as `distilled_guidance`.
For community edit adapters and other optional extensions, these are the base-model settings unless the source specifies otherwise.
| Recipe | CFG | Steps | Width × height | Source / notes |
| --- | --- | --- | --- | --- |
| `anima` | 4 | 20 | 1024 × 1024 | [Source](https://huggingface.co/circlestone-labs/Anima) The official ranges are 30–50 steps / CFG 4–5; CFG 4 selects the lower end. Edit uses the same base; tune the chosen editing LoRA separately. |
| `anima-edit` | 4 | 20 | 1024 × 1024 | [Source](https://huggingface.co/circlestone-labs/Anima) The official ranges are 30–50 steps / CFG 4–5; CFG 4 selects the lower end. Edit uses the same base; tune the chosen editing LoRA separately. |
| `animatediff-sd15` | 7.5 | 20 | 512 × 512 | [Source](https://raw.githubusercontent.com/guoyww/AnimateDiff/main/configs/prompts/1_animate/1_3_animate_ToonYou.yaml) |
| `boogu-image` | 4 | 20 | 1024 × 1024 | [Source](https://huggingface.co/Boogu/Boogu-Image-0.1-Base) |
| `boogu-image-edit` | 5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/Boogu/Boogu-Image-0.1-Edit) |
| `chroma` | 3 | 20 | 1024 × 1024 | [Source](https://huggingface.co/lodestones/Chroma1-HD/raw/main/README.md) |
| `chroma-radiance` | 3.5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/lodestones/Chroma1-Radiance/raw/main/inference_reference_code.py) |
| `cosxl-edit` | — | — | — | [Source](https://huggingface.co/stabilityai/cosxl) No numerical generation defaults in the author’s model card; previous test values were removed. |
| `ernie-image` | 4 | 20 | 1024 × 1024 | [Source](https://huggingface.co/baidu/ERNIE-Image) |
| `ernie-image-turbo` | 1 | 8 | 1024 × 1024 | [Source](https://huggingface.co/baidu/ERNIE-Image-Turbo) |
| `flex2` | 1 | 20 | 1024 × 1024 | [Source](https://huggingface.co/ostris/Flex.2-preview) |
| `flux1-canny` | 1 | 20 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.1-Canny-dev) |
| `flux1-depth` | 1 | 20 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.1-Depth-dev) |
| `flux1-dev` | 1 | 20 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.1-dev) |
| `flux1-fill` | 1 | 20 | 1232 × 1632 | [Source](https://huggingface.co/black-forest-labs/FLUX.1-Fill-dev) |
| `flux1-kontext` | 1 | 20 | — | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/flux/pipeline_flux_kontext.py) Output size normally follows the reference image; no fixed size is imposed by this recipe. |
| `flux1-pulid` | 1 | 20 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.1-dev) FLUX.1-dev defaults; identity conditioning is provided by PuLID. |
| `flux1-schnell` | 1 | 4 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.1-schnell) |
| `flux2-dev` | 1 | 20 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.2-dev) |
| `flux2-klein-4b` | 1 | 4 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B) |
| `flux2-klein-9b` | 1 | 4 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B) |
| `flux2-klein-base-4b` | 4 | 20 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B) |
| `flux2-klein-base-9b` | 4 | 20 | 1024 × 1024 | [Source](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B) |
| `flux3-klein` | 1 | 16 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/black-forest-labs/flux-3-image-klein/main/README.md) |
| `flux3-klein-base` | 4 | 20 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/black-forest-labs/flux-3-image-klein/main/src/flux3/pipeline.py) |
| `hidream-o1-dev` | 1 | 20 | 2048 × 2048 | [Source](https://huggingface.co/HiDream-ai/HiDream-O1-Image-Dev) |
| `hunyuan-video-15-i2v` | 6 | 20 | 1280 × 720 | [Source](https://huggingface.co/tencent/HunyuanVideo-1.5) 720p base I2V defaults; CFG/step-distilled and 480p checkpoints use different settings. |
| `hunyuan-video-15-t2v` | 6 | 20 | 1280 × 720 | [Source](https://huggingface.co/tencent/HunyuanVideo-1.5) 720p base T2V defaults; CFG-distilled / 480p checkpoints need their published settings. |
| `ideogram4` | 7 | 20 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/ideogram-oss/ideogram4/main/docs/inference.md) Constant-guidance pipeline defaults; this recipe does not reproduce the CLI’s per-step guidance schedule. |
| `instruct-pix2pix` | — | 10 | — | [Source](https://huggingface.co/timbrooks/instruct-pix2pix) Official quick-start example; output size follows the input image. |
| `iris-3b` | 3 | 20 | 1024 × 1024 | [Source](https://huggingface.co/speridlabs/iris-3b) |
| `krea2-raw` | 3.5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/krea/Krea-2-Raw) Base model defaults; edit requires a matching community LoRA with its own settings. |
| `krea2-raw-edit` | 3.5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/krea/Krea-2-Raw) Base model defaults; edit requires a matching community LoRA with its own settings. |
| `krea2-turbo` | 1 | 8 | 2048 × 2048 | [Source](https://huggingface.co/krea/Krea-2-Turbo) Official CLI example; edit requires a matching community LoRA. |
| `krea2-turbo-edit` | 1 | 8 | 2048 × 2048 | [Source](https://huggingface.co/krea/Krea-2-Turbo) Official CLI example; edit requires a matching community LoRA. |
| `lens` | 5 | 20 | 1440 × 1440 | [Source](https://huggingface.co/microsoft/Lens) Official model card: RL model, default base_resolution 1440. |
| `lens-turbo` | 1 | 4 | 1440 × 1440 | [Source](https://huggingface.co/microsoft/Lens-Turbo) Official model card: 4-step distilled model, default base_resolution 1440. |
| `lingbot-video-dense` | 3 | 20 | 832 × 480 | [Source](https://huggingface.co/robbyant/lingbot-video-dense-1.3b) Official first-stage inference example; optional refiner is separate. |
| `lingbot-video-moe` | 3 | 20 | 832 × 480 | [Source](https://huggingface.co/robbyant/lingbot-video-moe-30b-a3b) Official inference example. |
| `llada-image` | 5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/inclusionAI/LLaDA-Image) |
| `llada-image-turbo` | 1 | 4 | 1024 × 1024 | [Source](https://huggingface.co/inclusionAI/LLaDA-Image-Turbo) |
| `longcat-image` | 4 | 20 | 1344 × 768 | [Source](https://huggingface.co/meituan-longcat/LongCat-Image) |
| `longcat-image-edit` | 4.5 | 20 | — | [Source](https://huggingface.co/meituan-longcat/LongCat-Image-Edit) Reference-image-dependent output size; no fixed size is imposed. |
| `ltx23-dev` | 3 | 20 | 768 × 512 | [Source](https://raw.githubusercontent.com/Lightricks/LTX-2/main/packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py) Official single-stage parameters, before optional latent upscaling. |
| `ltx23-distilled` | 1 | 8 | 768 × 512 | [Source](https://raw.githubusercontent.com/Lightricks/LTX-2/main/packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py) Model card specifies 8 steps / CFG 1; size uses official pipeline’s first stage. |
| `ltx25-dev` | 3 | 20 | 768 × 512 | [Source](https://raw.githubusercontent.com/Lightricks/LTX-2/main/packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py) Official non-distilled pipeline parameters for model versions >= 2.4, before optional upscaling. |
| `mage-flow` | 5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/microsoft/Mage-Flow) |
| `mage-flow-edit` | 5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/microsoft/Mage-Flow) |
| `mage-flow-turbo` | 1 | 4 | 1024 × 1024 | [Source](https://huggingface.co/microsoft/Mage-Flow) |
| `ming-image` | 1 | 12 | 2048 × 2048 | [Source](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design) |
| `minimax-h3-fl2va` | 1 | — | 1344 × 768 | [Source](https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_h3) Guidance-distilled; official 768p canvas. No fixed step count is prescribed here; the upstream API counts sigma grid points rather than sd.cpp denoising steps. |
| `minimax-h3-ref2va` | 1 | — | 1344 × 768 | [Source](https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_h3) Guidance-distilled; official 768p canvas. No fixed step count is prescribed here; the upstream API counts sigma grid points rather than sd.cpp denoising steps. |
| `minit2i` | 2.5 | 20 | 512 × 512 | [Source](https://huggingface.co/MiniT2I/MiniT2I/raw/main/pipeline.py) B/16 example and native model size; L/16 uses CFG 6. |
| `moss-image-base` | 3 | 20 | 1024 × 1024 | [Source](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base) |
| `moss-image-sft` | 3 | 20 | 1024 × 1024 | [Source](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) |
| `nitrosd-realism` | 1 | 1 | 1024 × 1024 | [Source](https://huggingface.co/ChenDY/NitroFusion) |
| `ovis-image` | 5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/ATH-MaaS/Ovis-Image-7B) |
| `pid` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pid-15` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pid-15-flux2` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pid-15-qwen-image` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pid-15-sd3` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pid-flux2` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pid-qwen-image` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pid-sd3` | 1 | 4 | 2048 × 2048 | [Source](https://huggingface.co/nvidia/PiD) Standalone PiD decoding at 2K; guidance is handled by the upstream generator. |
| `pixart-alpha-1024` | 4.5 | 20 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py) |
| `pixart-alpha-512` | 4.5 | 20 | 512 × 512 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py) |
| `pixart-sigma` | 4.5 | 20 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py) |
| `qwen-image` | 4 | 20 | 1328 × 1328 | [Source](https://huggingface.co/Qwen/Qwen-Image) Official square resolution bucket; true_cfg_scale maps to cfg_scale. |
| `qwen-image-21` | 1 | 20 | 2048 × 2048 | [Source](https://raw.githubusercontent.com/QwenLM/Qwen-Image-2.1/main/README.md) Original checkpoint defaults; select the matching TAEQI2.1 to replace its VAE. |
| `qwen-image-21-turbo` | 1 | 8 | 2048 × 2048 | [Model card](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo), [fixed sigmas](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo/blob/main/model_index.json). Standalone Turbo weights; the fixed 8-step schedule is applied on model load. |
| `qwen-image-edit` | 4 | 20 | — | [Source](https://huggingface.co/Qwen/Qwen-Image-Edit) Official editing example; output dimensions follow reference-image aspect ratio. |
| `qwen-image-edit-2509` | 4 | 20 | — | [Source](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) Official editing example; output dimensions follow reference-image aspect ratio. |
| `qwen-image-edit-2511` | 4 | 20 | — | [Source](https://huggingface.co/Qwen/Qwen-Image-Edit-2511) Official editing example; output dimensions follow reference-image aspect ratio. |
| `qwen-image-layered` | 4 | 20 | 640 × 640 | [Source](https://huggingface.co/Qwen/Qwen-Image-Layered) Official recommended 640 resolution bucket; layered RGBA VAE is required. |
| `sd-turbo` | 1 | 1 | 512 × 512 | [Source](https://huggingface.co/stabilityai/sd-turbo) |
| `sd1-checkpoint` | 7.5 | 20 | 512 × 512 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) |
| `sd1-controlnet` | 7.5 | 20 | 512 × 512 | [Source](https://huggingface.co/lllyasviel/control_v11p_sd15_canny) |
| `sd1-inpainting` | 7.5 | 20 | 512 × 512 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) |
| `sd1-ip-adapter` | 7.5 | 20 | 512 × 512 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) SD 1.x base pipeline defaults; adapter strength is configured separately. |
| `sd2-512-checkpoint` | 7.5 | 20 | 512 × 512 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) |
| `sd2-768-checkpoint` | 7.5 | 20 | 768 × 768 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) |
| `sd2-inpainting` | 7.5 | 20 | 512 × 512 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) |
| `sd3-medium` | 7 | 20 | 1024 × 1024 | [Source](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) |
| `sd35-large` | 3.5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/stabilityai/stable-diffusion-3.5-large) |
| `sd35-medium` | 4.5 | 20 | 1024 × 1024 | [Source](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium) |
| `sdxl-checkpoint` | 5 | 20 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py) |
| `sdxl-controlnet` | 5 | 20 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py) SDXL pipeline defaults; compatible ControlNet weights are required. |
| `sdxl-inpainting` | 8 | 20 | 1024 × 1024 | [Source](https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1/raw/main/README.md) |
| `sdxl-ip-adapter` | 5 | 20 | 1024 × 1024 | [Source](https://raw.githubusercontent.com/huggingface/diffusers/main/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py) SDXL base pipeline defaults; adapter strength is configured separately. |
| `sdxl-photomaker` | — | — | 1024 × 1024 | [Source](https://huggingface.co/TencentARC/PhotoMaker) Use settings for the selected compatible SDXL checkpoint; no unique CFG/steps are prescribed by the model card. |
| `sdxl-turbo` | 1 | 1 | 512 × 512 | [Source](https://huggingface.co/stabilityai/sdxl-turbo) |
| `sdxs` | 1 | 1 | 512 × 512 | [Source](https://huggingface.co/IDKiro/sdxs-512-dreamshaper) |
| `sefi-image` | 4 | 20 | — | [Source](https://huggingface.co/SeFi-Image/SeFi-Image-5B-Base) Base/RL family defaults; Turbo checkpoints require 4 steps / CFG 1. |
| `sensenova-u15` | 4 | 20 | 2048 × 2048 | [Source](https://raw.githubusercontent.com/OpenSenseNova/SenseNova-U1/main/examples/README.md) |
| `wan21-flf2v` | 5 | 20 | 1280 × 720 | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.1/main/generate.py) |
| `wan21-i2v` | 5 | 20 | 832 × 480 | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.1/main/generate.py) 480p base defaults; also accepts compatible SkyReels V2 and single Wan 2.2 I2V experts. |
| `wan21-t2v` | 5 | 20 | 1280 × 720 | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.1/main/generate.py) 14B base defaults; single Wan 2.2 experts and acceleration LoRAs need matching settings. |
| `wan21-vace` | 5 | 20 | 832 × 480 | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.1/main/generate.py) |
| `wan22-i2v` | 3.5 | 20 | 1280 × 720 | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/wan/configs/wan_i2v_A14B.py) |
| `wan22-s2v` | 4.5 | 20 | — | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/wan/configs/wan_s2v_14B.py) The official S2V pipeline derives output dimensions from the input. |
| `wan22-t2v` | 3 | 20 | 1280 × 720 | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/wan/configs/wan_t2v_A14B.py) |
| `wan22-ti2v-5b` | 5 | 20 | 1280 × 704 | [Source](https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/wan/configs/wan_ti2v_5B.py) |
| `z-image` | 4 | 20 | 720 × 1280 | [Source](https://huggingface.co/Tongyi-MAI/Z-Image) |
| `z-image-l2p` | — | — | — | [Source](https://huggingface.co/zhen-nan/L2P) No numerical inference recommendations in the author’s model card; previous test values were removed. |
| `z-image-turbo` | 1 | 9 | 1024 × 1024 | [Source](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) |
TAE alternatives use the model-compatible weights listed by [TAESD](https://github.com/madebyollin/taesd) and [TAEHV](https://github.com/madebyollin/taehv).
TAE slots are omitted for pixel-space models, PiD, layered RGBA decoding and Mage-VAE paths that the current runtime does not map to a compatible TAE.
Reference in New Issue View Git Blame Copy Permalink