f2512_fp8 — Cinematic Realism & Emotional Depth
Description: This model is a high-fidelity FP8 merge optimized for 24GB VRAM cards (RTX 3090/4090), designed to deliver cinematic photorealism without compromising on emotional storytelling. Tested across diverse scenarios—from intimate portraits and complex lighting setups to humorous and narrative-driven scenes—f2512_fp8 consistently produces images with lifelike textures, accurate anatomy, and a distinct "fine art" aesthetic.
✨ Key Features:
Masterful Lighting: excels in natural light simulation, including Golden Hour glow, dramatic Chiaroscuro, and complex rim/backlighting. Handles subsurface scattering (SSS) on skin with high accuracy.
Rich Textures: Superior detail rendering in skin pores, vellus hair, wet/dry hair transitions, and diverse materials (fabric, fur, metal, wood).
Anatomy & Proportions: Reliable hand generation, expressive facial micro-expressions, and accurate handling of body diversity (contrast in age, size, and physique).
Versatility: seamlessly adapts between high-end fashion, realistic documentary style, and narrative humor/cinematic storytelling.
Optimized for 3090: The FP8 compression maintains quality while fitting comfortably into 24GB VRAM, allowing for 1024x1536 resolution generation without OOM errors.
⚙️ Recommended Settings:
Sampler:
Euler a(for soft, artistic results) orDPM++ 2M Karras(for sharper details).Steps:
30–40.CFG Scale:
5.5–6.5(Best balance of creativity and adherence).Scheduler:
KL_optimal.Resolution:
1024x1536or1024x1024.Clip Skip:
2(if applicable to the base architecture).
📸 What It Does Best:
Cinematic Portraits: Dramatic lighting, emotional depth, and photorealistic skin.
Texture Studies: Water droplets, wet hair, fabric folds, and organic surfaces.
Narrative Scenes: Complex interactions between subjects (e.g., human-animal bonding) and humorous/situational contexts.
📝 Notes:
For best results, use descriptive prompts focusing on lighting conditions and textures.
The model favors a "fine art photography" aesthetic; adding "photorealistic" or "cinematic" helps anchor the style.
FP8 version is a compressed variant that saves VRAM and generation time with negligible quality loss compared to FP16.
🔗 Merging Method: Merged using [Method, e.g., TIES-Merge / Linear Merge] with FP8 post-processing for efficiency.
Description
Finetuned, much more artfull
FAQ
Comments (35)
Hi! Whats the difference between base and early?
Early is raw, dirty model
But someone like it, base is more artfull
based on the description, probably the merge weights of the loras and model, also JIB has been updated so it's possible this uses that too
@llhappyll822 no, base merged with other unets - mine krea, srpo and chroma and more than 5 loras
I think its visible in images
These merges can be alright, I mean I think JIB mix is simply a merge of loras into the base. However it is troubling when you can look at a model and think it's another one because of the lack of diversity in the set. We should be training our own loras with real photos as any form of "fine tune", and I use that with a grimace. Otherwise it just feels like... masturbating. It isn't like I'm trying to blame you, or point you out. This was a problem with SDXL, too, you had to find some way of understanding what the author actually did to the model. And really the problem is our computational model as either a society, as the greed is starting to infect the pieces individually to a point where we're several years into an architecture cycle that was planned around this kind of obsolescence. And I specifically mean, I understand that our GPUs arent big enough to properly train these base models. But damnit, these lora stacks are getting troublesome. LoRa stands for Low-Rank Adaptation. We are only tickling the tips of the model in an attempt to change it. Things need to change.
Listen, this is a deeply philosophical question. I don't create anything—I just adapt the tool to my own ideas. I like to create what interests me, but I'm not ready to delve that deeply.
Sometimes I'm just too lazy to use a bunch of LORAs that I can't do without—I just merge them into a checkpoint—it's more convenient for me. I'm not Jack Ma, and I simply don't have the strength or resources to create and train a model from scratch.
@FASCIUM And there's a purpose for the merges, of course, especially if its a combination that works. Again, I wasn't trying to call you out or anything, just provide food for thought.
@FASCIUM My overarching point though is that you should be able to properly finetune the bases. These odd VRAM limits producers have placed on cards seems like the overall problem, so obviously we can't really be that upset that this is only viable outcome. Some model producers push in the right direction, but these other vulture-based tools are trading money for the huge machines necessary to run their monstrosities. I don't hate a modular enough approach, but this idea that we can't even represent certain things because the bases are impenetrable by virtue of their architecture is frustrating and flawed. I just don't think we should struggle so hard to represent our own humanity using tools that are only possible through vast theft of that essence. That we're just here to train more of their vast model shards that'll never be able to be run by any one person. I can appreciate that these developers have provided their architectures, I can appreciate that people like us can train these LoRas and make these merges so easily, but I simply don't like the direction we're going.
can u share settings of urs and workflow
I work FORGE, no workflows, settings opened in every image
@FASCIUM ok
I like how everyone these days just push dookie out and calls it a merge or trained model. This has the same face issue just like the jibs model, everyone just renders 5 images and post their model.
Testing? wtf is that.
@dannyds44 i like how everyone these days posts null information comments.
@jtk1996606 What kind of response do you expect? I don't hide the fact that my work is based on the works of the Masters, but I do what interests me. I go through trial and error. It's generally difficult to work with qwen; little comes of it, but I'm trying. Nevertheless, I do not tolerate rudeness and vulgarity like this author's.
@jtk1996606 The audience that is interested in these models is not observable through a microscope at all, we are all amateurs by and large. I started these projects just to illustrate my book of fairy tales, which I did, and then there are people who highly admire my work. I'm not a professional, I'm actually doing what I love in retirement, but I just like that I can save my work on the resource.
who need a face?
Hi, i may sound dumb but still... Can i use this model in stabel diffusion, like AUTOMATIC1111 Stable Diffusion WebUI?
Hello. A quick question. You mentioned merging SRPO and KREA, but this isn't possible since these are Flux models and therefore completely incompatible with the Qwen Image model architecture. How did you manage to do that?
Its possible, i do it ForgeNeo in checkpoint merger
Much harder to merge FLUX LORAs but its possibly too in old version kohya and old forge
Its hard way to try with many mistakes, but now you see- i found the way
Qwen and Flux use totally different architectures, so they aren't compatible. It's simply impossible to merge them, no matter what you think you did. It's like claiming you've merged Flux with SDXL.
@roberto_baggio here theory and practice diverge
Sure !
@roberto_baggio but you can just try to do it yourself - the best criterion of truth is practice
@FASCIUM Yeah, some people claim they can levitate with a lot of practice. Sadly, I don't have the time to waste trying it myself.
Please notify me when you release a paper explaining how to transfer or combine knowledge between different model architectures. (i.e, without using the distillation technique or training a new model from scratch)
@roberto_baggio empty discussion, my models (in dynamics) and works are available
I love your work, but that face is really burned in on this one. Every man and woman have the same face.
I know it's probably difficult to blur out the faces in your dataset, but it would be a HUGE step in taking your models to the extreme next level. Thank you for what you do!
You right, we have BIG BIG BIG trouble with train QWEN..... Kohya do not support of QWEN, only Comfy, but i do not work in Comfy
This probably come from the JibMix same face model used in this merge. I recognized the woman's face before I even saw that JibMix was included.