Model Reference

This document is the single reference for running inference on all supported models: capabilities, model directory layouts, run commands, and CLI options.

Before running any model, complete environment setup in docs/01-quick-start.md (Linux: docs/05-linux-guide.md).

--enable_compile: Not required for Hugging Face amd/ models — they ship pre-partitioned and ready to run. Use --enable_compile (-c) only on the first inference run with a locally built DD model (see docs/04-model-generate-guide.md).

Supported Models Summary
Common Arguments
SD 1.5 Family
SDXL Family
SD3 / SD3.5 Family
FLUX.1-schnell
FLUX.2 Klein
O1 Optimization Preset

Supported Models Summary

Model	Hugging Face ID	Resolution	T2I	I2I	ControlNet	OutPainting
SD 1.5	`amd/stable-diffusion-1.5-amdnpu`	512×512	✓
SD Turbo	`amd/sd-turbo-amdnpu`	512×512	✓
SD 2.1 (deprecated)	—	512/768	✓
SD 1.5 ControlNet	`amd/sd15-controlnet-amdnpu`	512×512	✓		Canny
SDXL-base	`amd/sdxl-base-amdnpu`	1024×1024	✓
SDXL-turbo	`amd/sdxl-turbo-amdnpu`	512×512	✓
Playground 2.5	`amd/playground-v2.5-1024px-aesthetic-amdnpu`	1024×1024	✓
SSD-1B	`amd/SSD-1B-amdnpu`	1024×1024	✓
DreamShaper XL	`amd/dreamshaper-xl-lightning-amdnpu`	1024×1024	✓
Segmind-Vega	`amd/segmind-vega-amdnpu`	1024×1024	✓	✓
SD3 Medium	`amd/stable-diffusion-3-medium-amdnpu`	512–1024	✓		Canny/Tile/Pose/Depth/Union	✓
SD3.5 Medium	`amd/stable-diffusion-3.5-medium-amdnpu`	512–1024	✓		Canny/Tile/Pose	✓
FLUX.1-schnell	`amd/FLUX.1-schnell-amdnpu`	288–1024	✓
FLUX.2 Klein	`amd/FLUX.2-klein-4B-amdnpu`	1024×1024	✓	✓

Dynamic resolution models support: 512×512, 512×768, 768×512, 576×1024, 1024×576, 768×1024, 1024×768, 1024×1024

All amd/ models are downloaded automatically from Hugging Face on first use when --model_path is omitted. Kolors requires a locally built DD model — see the Kolors section.

Common Arguments

These arguments apply to all run scripts:

Argument	Short	Default	Description
`--model_id`		required	Hugging Face model ID or model type identifier
`--model_path`		auto-download	Path to local model directory (auto-downloaded if omitted; see `RYZENAI_GENAI_SD_MODELS_ROOT`)
`--revision`		None	Hugging Face branch, tag, or commit hash (e.g. `"1.8.0"`)
`--prompt`		model default	Text description for image generation
`--prompt_2`		None	Second prompt for dual-encoder models (e.g. FLUX T5 encoder)
`--n_prompt`		`""`	Negative prompt (what to avoid)
`--width`	`-W`	model default	Output image width in pixels
`--height`	`-H`	model default	Output image height in pixels
`--num_inference_steps`	`-n`	model default	Number of denoising steps
`--guidance_scale`		model default	CFG scale (higher = more prompt adherence)
`--seed`		None	Random seed for reproducibility
`--t5_sequence_len`		model default	T5 sequence length (e.g. 77, 83, or 256 for Flux)
`--edit_image_path`		None	Reference image path(s) for FLUX.2 Klein image editing (accepts one or more paths)
`--output_path`		`generated_images`	Output directory for saved images
`--enable_compile`	`-c`	False	Compile DD fusion runtime on first run with a locally built DD model (not needed for HF `amd/` models)
`--enable_profile`	`-p`	False	Enable performance profiling
`--profiling_rounds`	`-pr`	4	Number of profiling rounds
`--no_excel`		False	Skip Excel metrics file; timing still printed to stdout
`--no_images`		False	Skip image saving (pipeline still runs; useful for pure benchmarking)
`--custom_op_path`		auto	`onnxruntime_providers_ryzenai.dll` / `libonnxruntime_providers_ryzenai.so` (auto-detected if omitted)
`--verbose`	`-v`	False	Enable verbose output
`--prompt_file_path`		None	Path to JSON prompt list file
`--num_images_per_prompt`		1	Number of images to generate per prompt
`-C` / `--controlnet`	`-C`	None	ControlNet mode: `Canny`, `Tile`, `Pose`, `Depth`, `Union`, `OutPainting`, `Removal`, `InPainting`, or `None`

Model Download Path

Models are cached under models/ in the project root by default. Override with RYZENAI_GENAI_SD_MODELS_ROOT:

# Windows
$env:RYZENAI_GENAI_SD_MODELS_ROOT = "D:\my_models"
python test/run.py --model_id "amd/sdxl-turbo-amdnpu" --prompt "..."

# Linux
export RYZENAI_GENAI_SD_MODELS_ROOT=/data/my_models
python test/run.py --model_id "amd/sdxl-turbo-amdnpu" --prompt "..."

Environment Variables (inference)

Variable	Windows	Linux
`RYZENAI_GENAI_SD_MODELS_ROOT`	Model download/cache directory	Same
`DD_PLUGINS_ROOT`	`set DD_PLUGINS_ROOT=.\lib\transaction\stx`	`export DD_PLUGINS_ROOT=./lib/transaction/stx`
Session setup (XRT, `LD_LIBRARY_PATH`)	See 01-quick-start.md	See 05-linux-guide.md Step 6

DD_PLUGINS_ROOT is required for SD3 / SD3.5 / FLUX dynamic-shape models.

SD 1.5 Family

Model Directory Layout

sd15/
├─ vae_decoder/
├─ unet/
├─ scheduler/
└─ tokenizer/

SD 1.5 Text-to-Image

# Auto-download from Hugging Face (recommended)
python test/run.py --model_id "amd/stable-diffusion-1.5-amdnpu" --prompt "a beautiful landscape"

# Local model path
python test/run.py --model_id "stable-diffusion-v1-5/stable-diffusion-v1-5" \
    --model_path path/to/sd15/models/ --prompt "a beautiful landscape"

Defaults: 512×512, 20 steps, guidance_scale=7.5

SD-Turbo Text-to-Image

# Batch size 1 (default)
python test/run.py --model_id "amd/sd-turbo-amdnpu" --prompt "a beautiful landscape"

# Batch size 2 (generates 2 images per prompt)
python test/run.py --model_id "amd/sd-turbo-amdnpu" --prompt "a beautiful landscape" \
    --num_images_per_prompt 2

Defaults: 512×512, 1 step, guidance_scale=0.0 (distilled)

SD 2.1 Text-to-Image

# SD 2.1-base (512×512)
python test/run.py --model_id "stabilityai/stable-diffusion-2-1-base" \
    --model_path path/to/sd2.1-base/models/

# SD 2.1 (768×768)
python test/run.py --model_id "stabilityai/stable-diffusion-2-1" \
    --model_path path/to/sd2.1/models/

Defaults: sd2.1-base: 512×512, 50 steps; sd2.1: 768×768, 50 steps

SD 1.5 ControlNet (Canny)

Generates images guided by an edge-detected control image.

Model directory:

sd15-controlnet/
├─ controlnet/
├─ vae_decoder/
├─ unet/
├─ scheduler/
└─ tokenizer/

python test/run.py \
    --model_id "amd/sd15-controlnet-amdnpu" \
    --prompt "a beautiful landscape" \
    --control_image_path path/to/canny/image.png

Additional ControlNet arguments:

Argument	Default	Description
`--control_image_path`	required	Path to the edge/control image
`--controlnet_conditioning_scale`	1.5	Strength of ControlNet guidance

Defaults: 512×512, 20 steps, guidance_scale=7.5

SDXL Family

Model Directory Layout

sdxl-turbo/
├─ vae_decoder/
├─ vae_encoder/
├─ unet/
├─ scheduler/
├─ tokenizer/
└─ tokenizer_2/

SDXL-base Text-to-Image

python test/run.py --model_id "amd/sdxl-base-amdnpu" --prompt "a beautiful landscape"

# Dynamic resolution (20 presets, 512×288 – 1024×1024)
python test/run.py --model_id "amd/sdxl-base-amdnpu" --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

Defaults: 1024×1024, 50 steps, guidance_scale=7.5

SDXL-turbo Text-to-Image

python test/run.py --model_id "amd/sdxl-turbo-amdnpu" --prompt "a beautiful landscape"

# Dynamic resolution (20 presets, 512×288 – 1024×1024)
python test/run.py --model_id "amd/sdxl-turbo-amdnpu" --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

Defaults: 512×512, 1 step (Runner default; orchestrator sdxl_turbo pipeline uses 4 steps), guidance_scale=0.0

Playground 2.5 Text-to-Image

python test/run.py --model_id "amd/playground-v2.5-1024px-aesthetic-amdnpu" \
    --prompt "a beautiful landscape"

# Dynamic resolution (20 presets, 512×288 – 1024×1024)
python test/run.py --model_id "amd/playground-v2.5-1024px-aesthetic-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

Defaults: 1024×1024, 50 steps, guidance_scale=3.0

SSD-1B Text-to-Image

SSD-1B is a compact SDXL variant with faster inference while maintaining 1024×1024 output quality.

python test/run.py --model_id "amd/SSD-1B-amdnpu" --prompt "a beautiful landscape"

Defaults: 1024×1024, 50 steps, guidance_scale=7.5

DreamShaper XL Text-to-Image

DreamShaper XL Lightning is a distilled SDXL model that generates high-quality images in very few steps.

python test/run.py --model_id "amd/dreamshaper-xl-lightning-amdnpu" \
    --prompt "a beautiful landscape"

# Dynamic resolution (20 presets, 512×288 – 1024×1024)
python test/run.py --model_id "amd/dreamshaper-xl-lightning-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

Defaults: 1024×1024, 6 steps (Lightning distilled; increase for higher quality), guidance_scale=7.5

Segmind-Vega Text-to-Image

python test/run.py --model_id "amd/segmind-vega-amdnpu" --prompt "a beautiful landscape"

Defaults: 1024×1024, 50 steps, guidance_scale=7.5

Segmind-Vega Image-to-Image

Transforms an existing image based on the prompt. The --strength parameter controls how much the original image is preserved vs. replaced.

python test/run.py --model_id "amd/segmind-vega-amdnpu" \
    --prompt "a beautiful landscape" \
    --control_image_path test/assets/controlimg_input_1024x1024.png \
    --strength 0.95

--strength reference:

0.3 (default) — 15 denoising steps; original image mostly preserved
0.7 — 35 steps; balanced mix of prompt and original
1.0 — 50 steps; original image mostly replaced by prompt

SD3 / SD3.5 Family

SD3 and SD3.5 use the MM-DiT (Multimodal Diffusion Transformer) architecture with three text encoders (2× CLIP + T5) and a 16-channel VAE.

Model Directory Layout

sd3/
├─ common/
│  ├─ scheduler/
│  ├─ tokenizer/
│  ├─ tokenizer_2/
│  ├─ tokenizer_3/
│  ├─ text_encoder/
│  ├─ text_encoder_2/
│  ├─ text_encoder_3_gptq_v2/
│  ├─ vae_decoder/
│  └─ vae_encoder/
└─ normal/
   ├─ transformer/
   ├─ controlnet-canny/   (ControlNet only)
   ├─ controlnet-pose/    (ControlNet only)
   └─ controlnet-tile/    (ControlNet only)

SD3 Text-to-Image

# 512×512
python test/run.py -C None --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape"

# 1024×1024
python test/run.py -C None --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" -H 1024 -W 1024

# With profiling
python test/run.py -C None --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" -p

Defaults: 512×512, 8 steps, guidance_scale=3.0

SD3.5 Text-to-Image

Note: The SD3.5 Hugging Face release (amd/stable-diffusion-3.5-medium-amdnpu) includes Text-to-Image (T2I) weights only. For ControlNet and painting modes, copy the relevant controlnet directories from your SD3 model folder — see the SD3.5 ControlNet and SD3.5 OutPainting / Removal / InPainting sections.

python test/run.py -C None --model_id "amd/stable-diffusion-3.5-medium-amdnpu" \
    --prompt "a beautiful landscape"

Defaults: 512×512, 8 steps, guidance_scale=4.5

SD3 ControlNet (Canny/Tile/Pose)

# Canny
python test/run.py -C Canny --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" --control_image_path path/to/image.png

# Tile
python test/run.py -C Tile --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" --control_image_path path/to/image.png

# Pose
python test/run.py -C Pose --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" --control_image_path path/to/image.png

ControlNet arguments:

Argument	Default	Description
`-C`	required	ControlNet type: `Canny`, `Tile`, `Pose`, or `None`
`--control_image_path`	required	Path to control/reference image
`--controlnet_conditioning_scale`	0.5	ControlNet guidance strength

SD3.5 ControlNet (Canny/Tile/Pose)

Note: The SD3.5 Hugging Face release does not include separate ControlNet weights. These modes reuse the SD3 controlnet weights — copy the relevant directories before running:
cp -r models/sd3/normal/controlnet-canny  models/sd3.5/normal/
cp -r models/sd3/normal/controlnet-tile   models/sd3.5/normal/
cp -r models/sd3/normal/controlnet-pose   models/sd3.5/normal/

Same commands as SD3 ControlNet, using stable-diffusion-3.5-medium-amdnpu:

python test/run.py -C Canny --model_id "amd/stable-diffusion-3.5-medium-amdnpu" \
    --prompt "a beautiful landscape" --control_image_path path/to/image.png

SD3 OutPainting / Removal / InPainting

Availability: Only the InPainting controlnet weights are published on Hugging Face (amd/stable-diffusion-3-medium-amdnpu). OutPainting and Removal weights are not currently available publicly.

Additional model directories:

sd3/
└─ normal/
   ├─ controlnet-outpainting/    # weights not on Hugging Face
   ├─ controlnet-removal/        # weights not on Hugging Face
   └─ transformer/
sd3/inpainting/
   ├─ controlnet-inpainting/
   └─ transformer/

# OutPainting — extend image borders (requires local weights)
python test/run.py -C OutPainting \
    --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    --control_image_path path/to/image.png \
    --image_pads 100 100 0 0    # [top, bottom, left, right] pixels

# Removal — remove objects from image (requires local weights)
python test/run.py -C Removal \
    --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "" \
    --control_image_path path/to/image.png \
    --control_mask_path path/to/mask.png

# InPainting — fill masked region guided by prompt (available on Hugging Face)
python test/run.py -C InPainting \
    --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "holiday decorations" \
    --control_image_path path/to/image.png \
    --control_mask_path path/to/mask.png

OutPainting/Removal/InPainting arguments:

Argument	Default	Description
`-C`	`OutPainting`	Mode: `OutPainting`, `Removal`, or `InPainting`
`--control_image_path`	required	Input image
`--control_mask_path`	required for Removal/InPainting	Binary mask (white = region to modify)
`--image_pads`	`[10,10,10,10]`	OutPainting border [top, bottom, left, right] in pixels
`--controlnet_conditioning_scale`	0.95 (OutPainting/Removal), 0.7 (InPainting)	ControlNet strength

Defaults: 1024×1024, 8 steps, guidance_scale=5.0

SD3.5 OutPainting / Removal / InPainting

Note: The SD3.5 Hugging Face release does not include separate ControlNet weights for these modes. These modes reuse the SD3 controlnet weights — copy the relevant directories from your SD3 model folder into the SD3.5 model directory:
cp -r models/sd3/normal/controlnet-inpainting  models/sd3.5/normal/
cp -r models/sd3/inpainting/                   models/sd3.5/

Same commands as SD3, using stable-diffusion-3.5-medium-amdnpu:

python test/run.py -C OutPainting \
    --model_id "amd/stable-diffusion-3.5-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    --control_image_path path/to/image.png

SD3 Dynamic Resolution

Dynamic resolution allows a single loaded model to generate images at multiple resolutions without reloading. Supported resolutions: 512×512, 512×768, 768×512, 576×1024, 1024×576, 768×1024, 1024×768, 1024×1024.

Dynamic model directory layout (note /dynamic/ subdirectory):

sd3/
├─ common/
│  ├─ vae_decoder/dynamic/
│  └─ vae_encoder/dynamic/
└─ normal/
   └─ transformer/dynamic/

# Text-to-image with dynamic resolution (20 resolutions)
python test/run.py -C None --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

# Specify resolution explicitly
python test/run.py -C None --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v -H 768 -W 1024

# With profiling
python test/run.py -C None --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v -p

Dynamic shape arguments:

Argument	Short	Default	Description
`--dynamic_shape`	`-ds`	False	Enable dynamic resolution
`--dynamic_shape_file_path`		`config/dynamic_shape_20x.json`	Custom dynamic shape config (20 presets by default)

SD3.5 Dynamic Resolution

python test/run.py -C None --model_id "amd/stable-diffusion-3.5-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

Note for SD3.5 with LoRA: Set --guidance_scale 0.0 explicitly when using LoRA weights.

SD3 ControlNet Dynamic Resolution

# SD3 ControlNet Canny with dynamic shapes
python test/run.py -C Canny --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" --control_image_path path/to/image.png \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

# SD3 ControlNet Union — supports Canny, Tile, Pose, Depth, Scribble
python test/run.py -C Union --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

Union ControlNet recommended settings by task:

Task	Steps	Control Scale	CFG
Canny	20	0.7	3.0
Pose	20	0.7	3.5
Tile	20	1.0	3.5
Depth	20	0.7	3.5
Scribble	20	0.7	3.5

# Union Canny example
python test/run.py -C Union --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v \
    --num_inference_steps 20 --controlnet_conditioning_scale 0.7 --guidance_scale 3.0 \
    --control_image_path ref/canny.jpg

SD3.5 ControlNet Dynamic Resolution

python test/run.py -C Canny --model_id "amd/stable-diffusion-3.5-medium-amdnpu" \
    --prompt "a beautiful landscape" --control_image_path path/to/image.png \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

FLUX.1-schnell

FLUX.1-schnell is a distilled diffusion transformer with dual text encoders (CLIP on iGPU + T5 on NPU) and a 16-channel VAE. It uses BFP16 quantization on the NPU for all major components.

Model Directory Layout

flux1-schnell/
├─ scheduler/
├─ tokenizer/                   # CLIP tokenizer
├─ tokenizer_2/                 # T5 tokenizer
├─ text_encoder/                # CLIP encoder (runs on iGPU via DirectML)
├─ text_encoder_3_gptq_v2/      # T5 encoder (runs on NPU)
├─ vae_decoder/dynamic/
└─ transformer/dynamic/

Running FLUX.1-schnell

# Run with dynamic resolution
python test/run.py --model_id "amd/FLUX.1-schnell-amdnpu" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v

# With custom resolution
python test/run.py --model_id "amd/FLUX.1-schnell-amdnpu" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v \
    -H 768 -W 1024

# With profiling
python test/run.py --model_id "amd/FLUX.1-schnell-amdnpu" \
    -ds --dynamic_shape_file_path config/dynamic_shape_20x.json -v -p

Defaults: 4 steps (distilled), guidance_scale=0.0 (ignored by FLUX.1-schnell)

FLUX-specific arguments:

Argument	Default	Description
`--prompt_2`	same as `--prompt`	Second prompt fed to the T5 encoder
`--dynamic_shape` / `-ds`	False	Enable dynamic resolution
`--dynamic_shape_file_path`	auto	Path to dynamic shape config JSON

Supported dynamic resolutions for FLUX transformer:

Resolution	latent_dim
288×512 / 512×288	576
384×512 / 512×384	768
512×512	1024
640×640	1600
768×768	2304
1024×1024	4096

FLUX.1-schnell is optimized for 4 steps. Using fewer steps yields diminishing quality gains. If you built the DD model locally, include --enable_compile on the first inference run.

FLUX.2 Klein

FLUX.2 Klein is a 4B-parameter diffusion transformer using the Qwen3 text encoder. It supports both text-to-image generation and image editing via --edit_image_path.

Model Directory Layout

flux2-klein/
├─ scheduler/
├─ tokenizer/
├─ text_encoder/                # Qwen3 encoder (runs on NPU)
├─ vae_decoder/
└─ transformer/

Running FLUX.2 Klein

# Text-to-image (recommended)
python test/run.py --model_id "amd/FLUX.2-klein-4B-amdnpu" \
    --prompt "a beautiful landscape" -v

# With profiling
python test/run.py --model_id "amd/FLUX.2-klein-4B-amdnpu" \
    --prompt "a beautiful landscape" -v -p

# Image editing — guide generation using a reference image
python test/run.py --model_id "amd/FLUX.2-klein-4B-amdnpu" \
    --prompt "turn the horse into a unicorn" \
    --edit_image_path path/to/input.png -v

Defaults: 1024×1024, 4 steps, guidance_scale=1.0, seed=0

FLUX.2 Klein–specific arguments:

Argument	Default	Description
`--edit_image_path`	None	Reference image for image editing (I2I). Can pass multiple paths.

O1 Optimization Preset

For SD3 and SD3.5 models, pass -O1 (or --optimize_o1) to enable a performance optimization preset that skips certain DiT steps during denoising.

python test/run.py -C None --model_id "amd/stable-diffusion-3-medium-amdnpu" \
    --prompt "a beautiful landscape" -O1

When O1 is beneficial: Use with num_inference_steps >= 8. O1 adds overhead that can dominate at very low step counts.

When O1 has no effect: For text-to-image (no ControlNet) with num_inference_steps <= 4, the runner automatically disables O1 to avoid degraded performance.

Output Files

Images: saved to generated_images/ (or --output_path) with names like stable-diffusion-v1-5_0_512x512_steps20_idx0_20260101_100000.png
Profiling data: .xlsx files in the same directory when --enable_profile is used
Result summaries: results/ directory when using scripts/run_pipelines.py

See docs/02-user-guide.md for batch pipeline configuration and execution modes.

FilesExpand file tree

03-model-reference-guide.md

Latest commit

History

03-model-reference-guide.md

File metadata and controls

Model Reference

Table of Contents

Supported Models Summary

Common Arguments

Model Download Path

Environment Variables (inference)

SD 1.5 Family

Model Directory Layout

SD 1.5 Text-to-Image

SD-Turbo Text-to-Image

SD 2.1 Text-to-Image

SD 1.5 ControlNet (Canny)

SDXL Family

Model Directory Layout

SDXL-base Text-to-Image

SDXL-turbo Text-to-Image

Playground 2.5 Text-to-Image

SSD-1B Text-to-Image

DreamShaper XL Text-to-Image

Segmind-Vega Text-to-Image

Segmind-Vega Image-to-Image

SD3 / SD3.5 Family

Model Directory Layout

SD3 Text-to-Image

SD3.5 Text-to-Image

SD3 ControlNet (Canny/Tile/Pose)

SD3.5 ControlNet (Canny/Tile/Pose)

SD3 OutPainting / Removal / InPainting

SD3.5 OutPainting / Removal / InPainting

SD3 Dynamic Resolution

SD3.5 Dynamic Resolution

SD3 ControlNet Dynamic Resolution

SD3.5 ControlNet Dynamic Resolution

FLUX.1-schnell

Model Directory Layout

Running FLUX.1-schnell

FLUX.2 Klein

Model Directory Layout

Running FLUX.2 Klein

O1 Optimization Preset

Output Files