Go to Extensions tab, Install from URL, use URL for this repo.
Important
For full GGUF compatibility, Forge must be up to date.
Otherwise you may get an error of the form RuntimeError: mat1 and mat2 shapes cannot be multiplied
Zero input images = normal inference (and don't need this extension).
One input image = slower inference, more VRAM needed.
Two input images = even slower inference, even more VRAM needed.
There are three input size/crop options:
- 'no change': might be useful for high resolution, but will require lots of VRAM and cause slower inference with large input images. Probably useful if you have already processed the input images to your own requirements: perhaps manually pasting multiple references into one larger image. Generally, do not use. Does not resize or crop.
- 'to output': the method used by previous versions. Resizes and centre-crops to match output resolution.
- 'to BFL recommended': resizes and centre-crops to BFL preferred resolutions (all around 1MP), matching aspect ratio as best as possible. Generally, use this.
If your generation size and input sizes match, and conform to recommended resolutions, all options will result in the same output.
reduce inputs to half width and height option shrinks the input images to half width and height - may reduce details but improves performance. Applies after the above option. If generating at low resolution and using 'to BFL recommended', also using this option is recommended.
Simply: encode input image(s) to base64; send the resultant strings to the extension.
Your payload should look like this, where ... is your normal specification of model, prompt, dimensions, steps, etc.:
{
... // usual payload
"alwayson_scripts": {
"Forge FluxKontext": {
"args": [
true, // enable
null, // image 1 encoded to base64 string
null, // image 2 encoded to base64 string
"to BFL recommended", // Kontext image size/crop, options are: "no change" | "to output" | "to BFL recommended"
false // reduce input to half width and height, after size/crop
]
}
}
}
For reference, the BlackForestLabs preferred resolutions are:
- 672 × 1568 and 1568 × 672
- 688 × 1504 and 1504 × 688
- 720 × 1456 and 1456 × 720
- 752 × 1392 and 1392 × 752
- 800 × 1328 and 1328 × 800
- 832 × 1248 and 1248 × 832
- 880 × 1184 and 1184 × 880
- 944 × 1104 and 1104 × 944
- 1024 × 1024
Prompting guide from BlackForestLabs.
Euler Simple is a reliable sampler/scheduler combo, others will work too and may give better results.
In img2img, you need very high denoise to see effect so that probably isn't a good way to increase preservation of the original. Quick tests show that inpainting works well.
Works with BlockCache. I've had good results with TeaCache, 3 uncached starting steps, 0.25 threshold, 1 max. consecutive uncached - reducing inference time by 40%. (I haven't tried to find an optimal speed/quality trade-off, so don't carve those parameters into stone and start a religion based on them do your own testing.) Ideal parameters will vary based on sampler, image size.
Compatibility with FluxTools Canny/Depth LoRAs is 0/10 - generation will fail. I can increase this to 1/10 (maybe 1/100 would be more accurate) with a padding based hack which allows generation to proceed, but there is minimal following of the control image and significant quality degradation. Increasing LoRa strength to force following the control image has increasingly severe quality cost.
