VERSION LINKS: FP8 • FP16 • NF4 • GGUF Q8_0 / Q6_K / Q5_KM / Q5_KS / Q5_0 / Q5_1 / Q4_KM / Q4_KS / Q4_0 / Q4_1 / Q3_KM / Q3_KS
V2 is an alternative to V1 with sharper good quality images from 4 steps. This version merges Schnell + Finetuned Dev + Hyper using the same but refined formula of variable block ratios from V1. Check comparison images below!
MAKE SURE TO RENAME YOUR FILES AFTER DOWNLOAD, CIVITAI GIVES THEM WRONG NAMES!
Tested sampler/scheduler for low steps:
ComfyUI: euler sampler, simple or beta scheduler.
Forge: euler, flux realistic sampler. KL Optimal or beta scheduler.
This model doesn't take guidance parameter, like schnell.
The versions with AIO (All in one) in the name include UNET + VAE + CLIP L + T5XXL (fp8). Also known as Checkpoint or Compact version.
Using BNB NF4 & GGUF quants in ComfyUI requires installing custom nodes that add special model loaders:
NF4 + Lora support: https://github.com/bananasss00/ComfyUI_bitsandbytes_NF4-Lora
(outdated) NF4 UNET: https://github.com/DenkingOfficial/ComfyUI_UNet_bitsandbytes_NF4
(outdated) NF4 AIO checkpoint: https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4
For using UNET versions, you also need to have the TEXT ENCODERS and VAE.
If you don't have them, download them from here:
T5XXL - CLIP L: https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main
GGUF T5XXL: https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf
VAE: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main/vae
Place the model in "models/diffusion_models" or "models/unet", both text encoders in "models/clip" and vae in "models/vae" folder.
In ComfyUI, use the standard flux workflow or add 'Load Diffusion Model', 'DualClipLoader' and 'Load VAE' nodes to replace the checkpoint loader and complete the setup.
In Forge, set the option "Diffusion in low bits" to "bnb-nf4"
Thanks to city96 for gguf quantization script.
Thanks to reddit user a_beautiful_rhind for bnb quantization script.










FLUX FUSION VERSION 1
Merge of Schnell and Dev variants of the Flux.1 model with a irregular smoothed ratio for each of the layers.
Quick comparison between versions. Prompts and settings at the end.
↓↓ Click show more for more examples and instructions ↓↓
Recommended use around 8 steps. If textures like skin look overworked, try lowering steps.
Comparison of V1 QUANTS:




Test parameters: 8 Steps, CFG 3.5, 1536x1536, seed 0
Prompts:
"Extreme closeup, frog face, star crystal structure, intricate designs, glowing hues. Extreme depth of field, celestial light, shimmering details, otherworldly charm, majestic elegance."
"extremely detailed 3d render portraits of a cyber-dragon themed flaming gothic arcane tech woman, cables, arcane tech-dragon inspired design, exposed machinery. the casing is glittery transparent tinted orange, red and black, allowing to see the internals. sophisticated fantasy design. abstract thematic background. extreme depth of field. dragon behind"
"classic pokemon 3DCG illustration. thick outlines. pokemon render style. extremely dynamic composition showcasing the special ability power. flowing pose with extreme closeup on the face in the upper area of the image. intense perspective, in motion movement effects, dynamic impactful vfx, eye catching.. A rock Pokémon with earthquake powers poses dynamically, its body a twisted mass of rugged terrain and molten lava flows. The upper area zooms in on its face, a mask of stone and fury with blazing eyes. Earth shatters beneath its feet as it stomps, unleashing seismic waves that ripple through the abstract background like a fractured canvas. Intense perspective compresses space, conveying unstoppable power. Vibrant colors dance: fiery oranges, electric blues, and smoldering grays. Movement effects blur edges, blurring boundaries between rock and energy. Impactful VFX burst forth in the foreground, echoing the Pokémon's raw force.. masterpiece, professional, best quality, sharp, extreme detail, Hyper-detailed, high-resolution, intricate, vivid. "
"Ethereal female face in 4K ultra closeup, eyes radiating eerie mystical aura with crystalline composition tinted purple-blue hues. Surrounding inferno blazes with dynamic flames and motion effects, creating a vertiginous atmosphere. Extreme depth of field emphasizes surreal otherworldly presence. Glowing eyes at the focal point contribute to haunting mystique, shot from an altered viewing angle emphasizing mysticism. Use Octane and Redshift raytracing for realistic fire and light effects, achieving ultra-realistic 3D render with intense, dreamlike quality."
"Ethereal star princess, diaphanous gown, shimmering stardust, intricate halo, luminous beauty. Night sky, glowing constellations, soft light, dreamy ambiance, mesmerizing allure."
"Sci-fi landscape, derelict alien structure, holographic iridescence, massive metal arches, dark skies, damaged antennas. Ground littered with debris, scattered wreckage, distant moon, dim light."
Description
Merge between Schnell and Dev variants of the Flux.1 model, with an irregular curve of ratios between blocks.
Recommended 8 steps from testing.
VAE and CLIP included in .safetensors file
FAQ
Comments (25)
1 example ...??? really ..common use ant or more unbelievable example ..f**this
Not sure if you looked in the wrong place... there are 6 examples for the fp8-e4m3fn version, the others aren't that much different so I only added one for those. I will add more when I have time.
@Anibaaal yes they are very misleading without mention that ef5 e4 thing whatever is that ..we don't all even understand that , explanation on ((((((5 pictures?!!!!))))) is not clear,
thank you.
@amazingbeauty sorry for the misunderstanding. I didn't add the samples in the other model versions and only put the examples in the version they are from. I thought it would be clear that way but I will add examples for the others too so it's complete and you can judge the results.
fp8 e5m2 and e4m3fn and fp16 are the same model just different "compression", fp8 use half the memory of fp16. They make slightly different images too.
empty account, no posts, comment is instantly aggressive. explains why I don't post content for people anymore.
@axicec do it for me my fren :P
@axicec you the one who really aggressive , you don't understand me at all , think about it your agro over me about 'how i talk' and you connects that with 'how you refuse to share your created content because of me! or some people you connected them fault to me!' , don't comment over my comment please cuz i will not reply you again ever. say what you want to say i'm out of any misunderstood talks.
sharing content is easy but sharing good content is hard enough to make or create, why i don't share content !? because i don't want to share shit.
for all ..why such 'empty account' as someone said, whole txt2img with help with paint (ai one not by hand) plus help with ipadapters...img2img is failed absolutely in making what want to prompt ~ make ..all the time in all situations.
Thanks for the merge, good for experimentation
Thanks for this merge I really like it. I tried the fp16 one with 8 steps and on some prompts it did a better job than the normal Dev model. Some prompts the Dev model didn't do what I wanted or forgot things I had in the prompt. I have to say tho that I am too impatient to run the Dev model with 40+ steps, so maybe it's because of that. I rather use this one with 8 steps.
did this 'the fp16 26gb' runs directly in comfy ? as loading as checkpoint ?
@amazingbeauty Yes, it does. You can use the Simple FP8 Checkpoint workflow from here: https://comfyanonymous.github.io/ComfyUI_examples/flux/
Just choose the fp16 smooth merge in the checkpoint loader and you can use it.
@LiaZ1 this is the best flux s .d merge that accept 4 to 5 steps out there ? or should i try others fp8 ones ? and thanks
You're welcome! The dev model is painfully slow for most setups, and the images composition and creativity aren't as interesting as Schnell in my opinion. I'm glad this worked for you
I tried you flux simple workflow with this models, but got error: "RuntimeError: ERROR: Could not detect model type of: E:\ComfyUI\models\checkpoints\Flux\fluxFusionDSSmoothMerge4_v0Fp8E4m3fn.safetensors" Are this models are really baked checkpoints or only diffusors?
It could be that you have an old version of comfy, I am using the same all in one models I uploaded and they work on my end with the simple workflow and others. Check if you have nodes like FluxGuidance or ModelSamplingFlux available in comfy, if not, you need to update it for flux support !
so the t5xxl is also fp8 ? baked in?
i think for my 24GB VRAM it's better to use fp8 flux with fp16 t5xxl
hi, yes, it’s baked with fp8 t5xxl, I will upload f16 versions later
you have merged them in ratio 1:1 schnell / dev, ye?
@tazztone Not 1:1, it's an irregular merge, almost all blocks have a different ratio from each other.
By the way, it will take a while to upload new versions with the fp16 encoder, but in the meantime you can download it from here just the t5xxl fp16 encoder and load it independently from the model with the 'DualCLIPLoader' node in comfy, you also need CLIP L. https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/text_encoders
Does this checkpoint go in unet or checkpoint loader? I'm using comfy ui
Hi, you should use the checkpoint loader, it includes clip and vae.
Update: for AIO models (all in one) use checkpoint loader, for UNET models use diffusion model loader.
Short update, still using it. And the SimpleTuner loras also work on this merge which is really great!
Glad you're enjoying it. That is amazing!!
Every time new Schnell + Dev merges are offered I try them, extensively. This "smooth" one is the only one that actually provides me with what I expect, additional quality with greater prompt adherence with respect to just "Schnell". It also has less chance of producing partial figures.
You may want to bump up to 6 steps but beyond that there is no discernable value that I was able to appreciate.
@Anibaaal - Thank you for taking the time to carefully craft this, it worked. The others being offered are failures at the time of this writing, 8/14/2024.
Thank you very much. You're right on the steps, I was using 8 for best quality but it really depends if you're doing art or photorealistic. More steps is not better and may even degrade the image.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.




