Selfora — realistic amateur checkpoint (Krea 2 Turbo, uncensored)
Selfora is a realism-first merge on Krea 2 Turbo built for one thing: believable amateur, shot-on-a-phone photos. Real skin with pores and texture, candid lighting, natural bodies — the kind of frame that looks like a selfie off someone's camera roll, not a glossy studio render. Ideal for AI-influencer / Instagram-style model content. Ideal for AI-influencer / Instagram-style model content. For that look with zero guesswork, pair it with my companion workflow — presets already tuned for this exact amateur-selfie output. → Selfism Optimal Krea 2 v.1. Every style LoRA is baked straight into the checkpoint — always on, nothing to stack or juggle. Fully uncensored, and just as strong on clean SFW. And it's Turbo, so you get there in a handful of steps.
Selfora started as a personal project — built for my own use and convenience, to get this exact look without stacking LoRAs every time. It came out well enough that I'm sharing it. Huge thanks to catlover1937 and to the author of the base model this checkpoint is built on, for kindly giving me permission to publish it.
Look & feel
Amateur / smartphone photoreal — real skin, real light, no plastic
Candid "phone camera" energy, not editorial studio gloss
Made for AI Instagram models & virtual influencers
Dialed-in out of the box — consistent without touching a single LoRA
Recommended settings (Krea 2 Turbo) — 2-pass is best
Two passes give the cleanest result: a base pass builds the image, a second partial-denoise pass refines skin and detail. Two variants — pick by whether you run RES4LYF. Both at CFG 1, 832×1216 (portrait) or 1024×1024.
⭐ Variant 1 — my setup (RES4LYF / ClownsharKSampler)
Pass 1 — base (ClownsharKSampler): sampler res_2s (exponential) · scheduler bong_tangent · 7 steps · eta 0.50 · bongmath on · denoise 1.0
optional extras: ClownOptions Detail Boost (weight 1.10 · model · soft · eta 0.45 · start_step 2), Sigmas Rescale (start 1.0 → end 0.0)
Pass 2 — refine (KSampler Advanced): euler · simple · 9 steps · start_at_step 4 → end 999 · add_noise on · return_with_leftover_noise off
Variant 2 — portable 2-pass (plain KSampler Advanced, no extra nodes)
Pass 1 — base: er_sde · simple · 8 steps · start 0 → end 8 · add_noise on · return_with_leftover_noise on
Pass 2 — refine: dpmpp_2m_sde · sgm_uniform · 9 steps · start_at_step 4 → end 999 · add_noise on · return_with_leftover_noise off
Quick single-pass: er_sde · simple · 8 steps · CFG 1 (fast, still good).
At CFG 1 the negative prompt does nothing — say what you do want, positively.
Prompting
No trigger words — the style lives in the weights. Just describe the scene.
Keywords for the best results with Selfora
The amateur / phone look is already baked in — you don't need any of these to get it. They just help you dial in a specific vibe (lighting, night, candid, camera) and get the most out of the model. Mix a couple from each group — you don't need all of them:
Amateur / candid base:
amateur iphone selfie,candid facial expression,casual everyday appearance,unposed,friend-circle snapshot quality,instagram story upload qualityiPhone / camera look:
shot on iPhone,iPhone 15 Pro photo,warm iphone color science,soft computational HDR,soft phone bokeh,slight sensor grain, front-camera selfie →wide-angle front-camera distortionReal skin (the realism boost):
natural unretouched skin texture,visible pores and freckles,minor blemishes,slight oily T-zone,natural makeup or bare face,realistic catchlights in eyesGrittier / night (optional):
low light phone photo,ISO 6400 sensor noise,slight motion blur,handheld,warm tungsten light,night mode processingCandid energy:
mid-laugh,genuine unposed expression,slightly tired eyes,casual
Avoid (these push it toward studio/AI-gloss): professional photography, studio lighting, DSLR, cinematic, retouched / airbrushed, 8k. At CFG 1 the negative prompt is ignored, so simply don't put these in the positive if you want the amateur look.
💡 Don't want to type any of this? Use the Selfism_Krea2_optimal_v1 workflow. All of the above — the exact amateur styling, lighting, camera looks and candid vibes — is already dialed in as ready one-click presets. It's tuned specifically for Selfora, so for this exact look with zero prompt fiddling, running Selfora inside Selfism is the easiest, most reliable way.
Content
Uncensored, NSFW and SFW both. It favors realistic amateur NSFW, but clean portraits and everyday shots come out just as well.
Best with my Selfism_Krea2_optimal_v1 workflow
Selfora runs best in my Selfism_Krea2_optimal_v1 workflow — it ships with lots of ready-made presets tuned for the selfie / amateur look and a full set of face-detailer presets, and the two sets are built to complement each other: the scene/selfie preset sets the shot, the face preset finishes the skin and detail — together they give a cohesive, consistent result. Powered by two custom nodes — Artfat LLM Prompter (image/text → prompt, on-device) and Artfat Face Preset (one-click phone-camera face presets). Load Selfora, pick your presets, shoot.
👉 Workflow: Selfism Optimal Krea 2 v.1 ( ONE NODE PROMPTER | FACE STYLER )
License & usage
Personal, non-commercial use only. Don't sell the model or anything made with it, and don't use it commercially. This mirrors the terms of moody's base model — please respect them. Don't re-upload or pass this checkpoint off as your own.
Credits
Built on moody Krea 2 Mix (uncensored) by moody — massive thanks for the base and for the OK to share this merge.
Baked LoRAs — credits
Full credit to the creators of the style LoRAs merged into this checkpoint (merged at my own balanced weights):
(Weights/balance are my own tuning and not published. If you made one of these and want the credit adjusted or removed, message me.)
Notes
fp8, self-contained one-file checkpoint — just load and go.
Planned for v2 (coming soon): an fp8_scaled build (quantized from the bf16 base for better quality) plus a full GGUF ladder for lower-VRAM setups. This v1 is the plain fp8 release.
Description
Selfora — GGUF Q4_K_M (runs on 8 GB VRAM)
Quantized to ~7 GB so it fits comfortably on 8 GB VRAM GPUs — no more OOM. Great for low-VRAM NVIDIA cards, laptops, and AMD.
⚙️ Load with the Unet Loader (GGUF) node (ComfyUI-GGUF) — not the regular checkpoint loader. Install ComfyUI-GGUF via Manager if you don't have it.
Recommended (Krea 2 Turbo): sampler er_sde · scheduler simple · 8 steps · CFG 1.
Smaller = slightly softer fine detail vs fp8/bf16, but full composition & likeness. For 12 GB+ use fp8; for RTX 30-series use the INT8 version.
FAQ
Comments (15)
you have to be number 1 in checkpoint list most downloaded what type of realism is this bruh i am in shocking state for each and every output this is crazy
Thank you, that honestly means a lot 🙏 I put a ton of work into the realism, so it's great to hear it hits. Fingers crossed it climbs the rankings — feel free to leave a review/like, that's what pushes it up there!
@Artfat What's the approximate percentage of young attractive males in the training data? Based on the sample images it looks heavily 1girl.
@ss9999 Most likely, the high percentage of young girls is due to the fact that my checkpoint is based on the “moody” model. For now, I’m addressing this age-related issue using the age slider in Krea2. This will need to be fixed in the next version.
@ss9999 We'll need to fix this in version 2
Is this a Turbo checkpoint model? I've tried it many times, and it feels like using the same prompt always generates very similar images, lacking that 'gacha' feel (randomness/variety). Does anyone know how to solve this?
Use dynamic sentences in beginning of prompt, details in the end. Try to use SeedVarianceEnhancer (firstable) or\and Krea2 Projector Tuner (BSS), play with parameters.
Could I ask which version of moody did you use as the base?
Which one is better quality wise fp8 or int8, in your humble opinion?
FP16 offers better image generation quality but is twice as large. I'll be uploading it to the website soon. Int8 and FP8 offer the same quality. Int8 is optimized for use with the 3000 series of graphics cards. FP8 is optimized and runs faster with the 4000 series of graphics cards.
I use INT8 with 5060Ti 16go, much faster than FP8. I'm missing something ?
@liquidsnake4ch287 You're not missing anything. In most ComfyUI paths fp8 is really just a storage format — the weights get dequantized back to bf16 for the actual matmul, so you save VRAM but not compute time. INT8 can hit a real accelerated integer kernel instead, which is why it comes out faster for you. Native fp8 math only pays off if the pipeline actually uses fp8 matmul, not just fp8 weights.
@Artfat I search what is matmul, I just added --fast to my bat. I will test on 4 run. I came back.
But many thanks for your checkpoint, I really like it.
@Artfat After many tests, optimisations, etc. Now FP8 is faster than before. BUT. INT8 is always faster. THX for help.
The issue with using moody as a base is that it tends to cause facial drift with character LoRAs, resulting in features like small round faces and big eyes.