generation
Run and troubleshoot sd-scripts image generation and minimal inference CLIs, including txt2img, img2img, inpainting, prompt files, SDXL conditioning, LoRA/network application, ControlNet/LLLite, and model-family inference scripts.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- generation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Run and troubleshoot sd-scripts image generation and minimal inference CLIs, including txt2img, img2img, inpainting, prompt files, SDXL conditioning, LoRA/network application, ControlNet/LLLite, and model-family inference scripts.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/commands.md9.0 KB
- 参考文档references/prompt-formats.md5.1 KB
- 参考文档references/troubleshooting.md5.7 KB
- 脚本scripts/validate_prompt_file.py7.3 KB
03 SKILL.md 正文
sd-scripts Generation
Use this sub-skill when the user wants to generate images, validate prompt files, apply trained LoRA weights during inference, or diagnose generation failures. For LoRA training route to ../training; for model merging/conversion route to ../model-utilities; for dataset TOML route to ../data-preparation.
Choose the Entry Point
- Use
gen_img.pyfor SD 1.x, SD 2.x, SDXL, txt2img, img2img, inpainting-by-mask, prompt files, highres fix, LoRA networks, ControlNet, and ControlNet-LLLite. - Use
sdxl_gen_img.pywhen you want the SDXL-focused script surface and the same prompt-file conventions. - Use
gen_img_diffusers.pywhen the model is a Diffusers pipeline/folder and the user wants Diffusers-style loading. - Use
inpainting_minimal_inference.pyfor a small inpainting script with--ckpt_path,--image,--mask, and--sdxlrather than the broadergen_img.pyinterface. - Use family minimal inference scripts for newer architectures:
sdxl_minimal_inference.py,sd3_minimal_inference.py,flux_minimal_inference.py,lumina_minimal_inference.py,hunyuan_image_minimal_inference.py,anima_minimal_inference.py, andanima_minimal_inference_control_net_lllite.py.
Fast Start Commands
Keep commands model-path agnostic and replace placeholders with the user’s files:
python gen_img.py --ckpt <model.ckpt-or-safetensors-or-diffusers-id> --outdir outputs \
--fp16 --xformers --W 512 --H 768 --steps 30 --sampler k_euler_a \
--prompt "masterpiece, detailed illustration --n low quality, blurry"
python gen_img.py --ckpt <model> --outdir outputs --bf16 --sdpa --no_preview \
--from_file prompts.txt --images_per_prompt 4 --batch_size 2 --seed 1234
python gen_img.py --ckpt <sdxl-model> --sdxl --outdir outputs --bf16 --sdpa \
--original_width 1024 --original_height 1024 --target_width 1024 --target_height 1024 \
--prompt "cinematic studio portrait --n distorted face"
Before running a batch, validate prompt-file syntax without loading models:
python skills/sd-scripts/sub-skills/generation/scripts/validate_prompt_file.py prompts.txt --family gen-img
Required References
references/commands.md: command templates for txt2img, img2img, inpainting, SDXL, LoRA, ControlNet, LLLite, and minimal inference scripts.references/prompt-formats.md: prompt-file syntax, negative prompts, prompt weights, dynamic prompts, per-line options, and model-family differences.references/troubleshooting.md: diagnosis for model-family flags, path errors, headless preview, precision artifacts, xformers/SDPA, OOM, LoRA count mismatch, ControlNet mismatches, and offline tokenizer cache.
Practical Workflow
- Identify the model family first: SD1/SD2/SDXL uses
gen_img.py; SD3/Flux/Lumina/Hunyuan/Anima use family minimal inference scripts. - Confirm required model components exist: checkpoint or DiT plus text encoders plus VAE/AE as required by the selected script.
- Select precision and memory flags conservatively: start with
--fp16or--bf16, reduce--batch_size, reduce resolution, then add VAE slicing/offload flags if needed. - Validate prompt files with the bundled script; it catches malformed per-line options, missing values, invalid numeric fields, and likely LoRA multiplier count mistakes.
- If applying LoRA at inference, match each
--network_modulewith a--network_weightsentry forgen_img.py, or use the family script’s--lora_weights/--lora_weightconvention. - For ControlNet/LLLite, verify image, mask, and control dimensions and ensure ControlNet and LLLite are not mixed in one
gen_img.pycommand.
Common Decisions
- Prefer
--no_previewon servers, containers, CI, SSH sessions, and any headless environment. - Prefer
--sdpaon modern PyTorch when xformers is unavailable; use--xformersonly when the package is installed and compatible. - For SD2 v-parameterized checkpoints, include both
--v2and--v_parameterization; missing--v_parameterizationcan produce brown images. - For black images or NaNs, retry with
--bf16,--no_half_vae, lower resolution, or fp32 depending on hardware support. - For prompt weighting, use WebUI-like
(term:1.2),(term), and[term]; keep generation options separated with spaces, such asprompt text --n bad anatomy --w 768.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).