-
-
Notifications
You must be signed in to change notification settings - Fork 426
FLUX
FLUX.1 family consists of 3 variations:
-
Pro
Model weights are NOT released, model is available only via Black Forest Labs -
Dev
Open-weight, guidance-distilled from Pro variation, available for non-commercial applications -
Schnell
Open-weight, timestep-distilled from Dev variation, available under Apache2.0 license
Additionally SD.Next includes pre-quantized variations of FLUX.1 Dev variation: qint8
, qint4
and nf4
Pick variant that uses less memory as model in original form has very high requirements
Important
Allow gated access
This is a gated model, you need to accept the terms and conditions to use it
For more information see Gated Access Wiki
Important
Set offloading
Set appropriate offloading setting before loading the model to avoid out-of-memory errors
For more information see Offloading Wiki
Important
Choose quantization
Check compatibility of different quantizations with your platform and GPU!
For more information see Quantization Wiki
Tip
Use reference models
Use of reference models is recommended over manually downloaded models!
Simply select it from Networks -> Models -> Reference
and model will be auto-downloaded on first use
Important
Do not attempt to assemble a full model by loading all individual components
That may be how some other apps are designed to work, but its not how SD.Next works
Always load full model and then replace individual components as needed
Warning
If you're getting error message during model load: file=xxx is not a complete model
It means exactly that - you're trying to load a model component instead of full model
FLUX.1 model consists of:
- Unet/Transformer: MMDiT
- Text encoder 1: CLIP-ViT/L,
- Text encoder 2: T5-XXL Version 1.1
- VAE
When using reference models, all components will be loaded as needed.
If using manually downloaded model, you need to ensure that all components are correctly configured and available.
Note that majority of available downloads are not actually all-in-one models and are instead just a part of the full model with individual components.
Tip
For convience, you can add setting that allow quick replacements of model components
to your quicksettings by adding
Settings -> User Interface -> Quicksettings list -> sd_model_checkpoint, sd_unet, sd_vae, sd_text_encoder
There are already many FLUX.1 unofficial variations available
Any Diffuser-based variation can be downloaded and loaded into SD.Next using Models -> Huggingface -> Download
For example, interesting variation is a merge of Dev and Schnell variations by sayakpaul: sayakpaul/FLUX.1-merged
SD.Next includes support for FLUX.1 LoRAs
Since LoRA keys vary significantly between tools used to train LoRA as well as LoRA types,
support for additional LoRAs will be added as needed - please report any non-functional LoRAs!
Also note that compatibility of LoRA depends on the quantization type! If you have issues loading LoRA, try switching your FLUX.1 base model to different quantization type
Typical all-in-one safetensors file is over 20GB in size and contains full model with transformer, both text-encoders and VAE
Since text encoders and VAE are same between all FLUX.1 models, using all-in-one safetensors is not recommended due to large duplication of data
Unet/Transformer component of FLUX.1 is a typical model fine-tune and is around 11GB in size
To load a Unet/Transformer safetensors file:
- Download
safetensors
orgguf
file from desired source and place it inmodels/UNET
folder
example: FastFlux Unchained - Load FLUX.1 model as usual and then
- Replace transformer with one in desired safetensors file using:
Settings -> Execution & Models -> UNet
SD.Next allows changing optional text encoder on-the-fly
Go to Settings -> Models -> Text encoder and select the desired text encoder
T5 enhances text rendering and some details, but its otherwise very lightly used and optional
Loading lighter T5 will greatly decrease model resource usage, but may not be compatible with all offloading modes
Tip
To use prompt attention syntax with FLUX.1, set
Settings -> Execution -> Prompt attention to xhinker
Example image with different encoder quantization options
SD.Next allows changing VAE model used by FLUX.1 on-the-fly
There are no alternative VAE models released, so this setting is mostly for future use
Tip
To enable image previews during generate, set Settings -> Live Preview -> Method to TAESD
To further speed up generation, you can disable "full quality" which triggers use of TAESD instead of full VAE to decode final image
FLUX.1 at the moment supports only Euler FlowMatch scheduler, additional schedulers will be added in the future
Due to specifics of flow-matching methods, number of steps also has strong influence on the image composition, not just on the way how its resolved
Example image at different steps
Additionally, sampler can be tuned with shift parameter which roughly modifies how long does model spend on composition vs actual diffusion
Example image with different sampler shift values
Support for all InstantX/Shakker-Labs models including Union-Pro
FLUX.1 ControlNets are large at over 6GB on top of already very large FLUX.1 model
as such, you may need to use offloading:sequential which is not as fast, but uses far less memory
When using union model, you must also select control mode in the control unit
Performance and memory usage of different FLUX.1 variations:
dtype | time (sec) | performance | memory | offload | note |
---|---|---|---|---|---|
bf16 | >32 GB | none | *1 | ||
bf16 | 50.47 | 0.40 it/s | balanced | *2 | |
bf16 | 94.28 | 0.21 it/s | 1.89 GB | sequential | |
nf4 | 14.69 | 1.36 it/s | 17.92 GB | none | |
nf4 | 21.02 | 0.95 it/s | balanced | *2 | |
nf4 | sequential | err | |||
qint8 | 15.42 | 1.30 it/s | 18.85 GB | none | |
qint8 | balanced | err | |||
qint8 | sequential | err | |||
qint4 | 18.37 | 1.09 it/s | 11.38 GB | none | |
qint4 | balanced | err | |||
qint4 | sequential | err |
Notes:
- *1: Memory usage exceeeds 32GB and is not recommended
- *2: Balanced offload VRAM usage is not included since it depends on desired threshold
© SD.Next