TheNoise is an open-source image generation / editing engine made specifically to run well on Strix Halo (gfx1151) and other ROCm-capable AMD iGPUs and dGPUs. It is tuned to perform extremely well on the machine it runs on.

TheNoise loads one model at a time and generates images from text prompts. Editing-capable models - like Qwen-Image 2.1, Qwen-Image-Edit, FLUX.2 Klein, Mage-Flow, Flux.3 and Fibo 2.0 - can also edit an existing image from a text instruction (image + prompt → edited image).

TheNoise can be used standalone, from the command line or through a webui or through an OpenAI-compatible server like Lemonade, with which it is already integrated.

Main

Edit

Upscale

Control Maps

---

Features

TheNoise ships:

• image generation / editing support for major open-weights models,
• Lemonade Server integration - TheNoise ships as its thenoise backend, see Using TheNoise with Lemonade Server,
• a built-in 2× refiner-based (SesquiLSR) upscaler - fast and high-quality upscaling without loading extra model files,
• pixel-space upscalers (Real-ESRGAN) up to 4× - standard ESRGAN-based models,
• post-processing in the decode tail - colour grading (saturation, vibrance and a white-balance pair), RCAS sharpening and film grain,
• control-map extraction for conditioning workflows - canny edges, line-art and soft-edge maps, depth maps and openpose-style pose skeletons,
• LoRA support - one or more LoRAs per image, each with its own weight.

Supported models

Model	Generate	Edit	Details
Anima, small and fast	✓	—	anima
Krea 2, highest image quality	✓	—	krea2
Qwen-Image 2.1, generation and editing in one model	✓	✓	qwen-image-2.1
MOSS-Image, joint-stream DiT with a Qwen3.5 conditioner	✓	—	moss-image
Z-Image / Z-Image-Turbo, quality at 8 steps	✓	—	zimage
Flux.2 Klein 4B / 9B, 4 steps, with editing	✓	✓	flux2-klein
Qwen-Image / Qwen-Image-Edit, generation and editing	✓	✓	qwen-image
Ming-Image 0.1, RGBA output with real transparency	✓	—	ming-image
Mage-Flow / Turbo, the fastest model, with editing	✓	✓	mage-flow
Flux.3, 15 steps with editing	✓	✓	flux3
Fibo 2.0, Bria's 4-step generate-and-edit model	✓	✓	fibo2

New models are added over time. PRs adding model support are welcome.

How does it compare to ComfyUI?

ComfyUI is a general-purpose, node-based framework and remains the better choice for advanced, customizable workflows. TheNoise is a focused engine, and it is a good fit when:

• you are running a Strix Halo and would like to start generating images quickly, without having to care about "workflows",
• you prefer a simple command line or a small UI over building and maintaining workflows,
• you want a small, stable image-generation endpoint that other software can call,
• you would rather have an engine optimized for your hardware than a general-purpose one.

Performance

TheNoise is tuned for the hardware it runs on, and its performance is quite similar to ComfyUI's - and sometimes even better. The numbers below are seconds per image on a Strix Halo (gfx1151, 128 GB unified), measured after a warmup run.

Text to image (generation):

Model & settings	1024×1024	1536x2048
Krea 2 Turbo · BF16 · 8 steps	33.8s	117.6s
Krea 2 Turbo · INT8-ConvRot · 8 steps	26.7s	99s
Anima Base · 20 steps · CGF 3	29.7s	111s
Anima Turbo · 8 steps	6.6s	25.1s
Mage-Flow Turbo · BF16 · 4 steps	2.4s	8.4s
Z-Image Turbo · 8 steps	14.3s	53.8s
Flux.2 Klein 9B · INT8-ConvRot · 4 steps	9.6s	34.3s
Qwen-Image 2512 · BF16 · 4 steps	9.7s	34.9s
Qwen-Image 2.1 Turbo · INT8-ConvRot · 8 steps	17.8s	65s
Ming-Image 0.1 · INT8-ConvRot · 12 steps	21s	84.8s
Flux 3 Klein · INT8-ConvRot · 15 steps	13.4s	75.8s
MOSS-Image · BF16 + 8 step DMD LoRA · 8 steps	12.6s	51.2s
Fibo 2.0 Turbo Edit · INT8-ConvRot · 4 steps	9.2s	35.6s

Image + simple instruction to edited image (editing):

Model & settings	1024×1024 · KV-cache OFF	1024×1024 · KV-cache ON
Flux.2 Klein 9B · INT8-ConvRot · 4 steps	19.2s	13.4s
Qwen-Image-Edit 2511 · BF16 · 4 steps	21.7s	14.2s
Qwen-Image 2.1 Turbo · INT8-ConvRot · 8 steps	—	21.3s
Mage-Flow Turbo · BF16 · 4 steps	4.9s	— (no KV cache)
Flux 3 Klein · INT8-Convrot · 15 steps	41.8s	— (no KV cache)
Fibo 2.0 Turbo Edit · INT8-ConvRot · 4 steps	15.5s	20.6s

TheNoise 0.11.0, Strix Halo (gfx1151, 128 GB unified)

Quick start

Install TheNoise and generate your first image in a few commands.

uv is the only prerequisite - it provides the Python interpreter and installs every dependency. If you don't have it already do:

curl -LsSf https://astral.sh/uv/install.sh | sh
source ~/.bashrc

then proceed with TheNoise installation

# 1. clone the repo
git clone https://github.com/lemonade-sdk/thenoise.git
cd thenoise

# 2. bootstrap the environment
./thenoise.sh --help

# 3. download a model
uv pip install -e ".[scripts]"
.venv/bin/python scripts/download.py --model anima

# 4. generate
./thenoise.sh generate \
  --dit ./models/anima/split_files/diffusion_models/anima-turbo-v1.0.safetensors \
  --vae ./models/anima/split_files/vae/qwen_image_vae.safetensors \
  --text-encoder ./models/anima/split_files/text_encoders/qwen_3_06b_base.safetensors \
  --prompt "a fox walking in the snow" --out fox.png

Using TheNoise

If you want to…	Use
generate an image from the command line	the CLI - generate, edit, upscale, controlmap
work through a browser	the web UI at http://localhost:8000/ when running serve
call it from other software	the HTTP API - /text2image, /edit, /upscale, /controlmap
extract a control map from an image	controlmap - the CLI, the /controlmap endpoint or the web UI's Control Maps tab

Documentation

Setup	from a released build to your first image
CLI reference	all generate / edit / serve / upscale flags, upscaling, LoRAs
HTTP API	endpoints, request/response reference, curl and Python examples
Development & Contribution	building from source, tests, portable builds, adding models

Acknowledgments

This project incorporates code from:

1. Musubi Tuner
2. SD Scripts
3. SesquiLSR

plus smaller snippets from other sources or transitively inherited through the above codebases.