I am not the original author of this model. This is a repost from HuggingFace project page here:<a target="_blank" rel="ugc" href="https://huggingface.co/city96/Flux.1-Heavy-17B">https://huggingface.co/city96/Flux.1-Heavy-17B</a> <h3 id="about-8h3fc3wok">About</h3>This is a 17B self-merge of the original 12B parameter <a target="_blank" rel="ugc" href="https://huggingface.co/black-forest-labs/FLUX.1-dev">Flux.1-dev</a> model.Merging was done similarly to 70B-&gt;120B LLM merges, with the layers repeated and interwoven in groups.<pre><code>Final model stats:
 p layers: [ 32]
 s layers: [ 44]
 n params: [17.17B]</code></pre><h3 id="comfyui-ctkrq5z3i">ComfyUI</h3>Just load it normally via the "Load Diffusion Model" node and placed in the "diffusion_models" folder. For the original 35GB file you need like 80GBs of system RAM on windows for it to not swap to disk lol. It requires about 35-40GBs of VRAM for inference, assuming you offload the text encoder and unload it during VAE decoding. Partial offloading works if you have enough system RAM.<h2 id="settings-lora-compatibility-yxksym68k">Settings? LoRA compatibility?</h2>Just use the same settings you'd use for regular flux. LoRAs do seem to have at least some effect, but the blocks don't line up so don't expect them to work amazingly.Important Notes From Uploader (me)<ul><li>This rendered a 512x512px image, 30 steps, takes my 8GB VRAM GPU 80 seconds to render.</li><li>I took the 35GB file and quantized it down to q_4 creating the 9.6GB file.</li><li>I will also be publishing q_2 and q_8 versions to compare.</li></ul>

Flux Heavy 17B

This must go into the diffusion_models folder, you must use the Load Diffusion Models node in ComfyUI

Files with SHA256: 6d0b6e827afb6557be350fd8fe0d89e95114bc37b9f09480a99671d42c5175e6