### 💬 Questions or setup issues?
Join the [Artfat AI Discord](https://discord.gg/r9tepyNSw) — I help with installation, workflows and prompting there, and it's where the community hangs out.
Selfism AIO Krea 2 v.1 — the workflow behind Selfora
My signature preset-driven pipeline for realistic amateur / phone-photo generation. Built around my checkpoint model Selfora but works with any Krea 2 model. The AIO version adds Edit, ControlNet and Colorgrade modes on top of the standard generation.
Solo project — the free stuff stays free ☕ Want new models early + a Discord role? Patreon's open, totally optional → https://www.patreon.com/c/ArtfatAI
What's inside
Artfat LLM Prompter — image/text → prompt, on-device (no API). Dozens of ready selfie/amateur style presets.
Face Style Preset — one-click phone-camera face presets for FaceDetailer (skin, grain, the "real" look).
2-pass gen (small base → ×1.9 → ×2 upscale), FaceDetailer, skin-contrast pass.
Edit Mode — Krea 2 Identity Edit: restage, change clothes, outfit transfer, masked local edits.
ControlNet Mode — reproduce the pose & composition of a reference image.
Colorgrade (LUT) — final-look grading, film grain, and a standalone image editor.
Resolution reference table baked in — pick a ratio, go.
How to use (standard generation)
Load Selfora (or any Krea 2 checkpoint).
Pick a preset in the LLM Prompter, drop a reference image (optional).
Choose resolution from the table → Run.
No trigger words. Sampler + settings pre-dialed.
✏️ EDIT MODE — Krea 2 Identity Edit
Instruction-based editing that keeps the subject's identity. Runs on the Turbo checkpoint.
LoRAs
krea2_identity_edit_v1_2 — general edit LoRA, handles all operations below.
krea_outfittransfer — dedicated outfit / clothing swap; trigger "transfer the outfit".
What it can do
Re-stage a person — keep the face, change scene/light: "Create a photo of this person at a night market"
Change clothing — "Change her black dress to a red dress"
Add / remove / replace — "Remove the necklace" / "Replace the woman with a big orangutan"
Outfit Transfer — apply an outfit from a reference: "replace the clothes of the woman in the image 1, transfer the outfit from the image 2"
Restyle whole image — "make it look like 90s film photography"
2-image compose — scene = Image 1, person = Image 2 (order is fixed)
Turn Edit Mode ON (Fast Groups Bypasser)
Load VAE for Edit Mode → krea2RealVae_v10 or Wan2.1_VAE_upscale2x_imageonly_real_v1 (the latter also gives a ×2 upscale).
Enable reference image 1 (and image 2 if used).
ENABLE Krea2 identity edit.
DISABLE the first sampler ClownShark — required.
Two references → enable both ref toggles + Enable edit Image 2.
Mask edit (first image only): right-click the reference → Open in Mask Editor → paint the area → Save → ENABLE Enable Mask edit.
💡 Stronger face identity: in the Krea2 identity edit subgraph set ref_boost = 4 (locks reference face / hair / clothes). Lower it if a bigger edit isn't "taking".
Settings: Turbo — 8–12 steps · CFG 1.0 · LoRA 1.0 · image ≤ 2 MP, dimensions divisible by 8.
Prompting: plain natural-language instructions, no trigger word — except Outfit Transfer.
🎯 CONTROLNET MODE — Krea 2 (Depth)
Reproduce the pose & composition of a reference image via a Control-LoRA (it patches the model, not a classic ControlNet). ⚠️ Output is a similar image built from the LLM's reading of the input — not an exact copy.
How to use
Enable Enable Image 1 → load the image whose pose / composition you want to reproduce.
Flip ON CONTROL NET (FIRST REF IMAGE ON) → true — routes the source latent into the ControlNet Group. (required)
In Fast Groups Bypasser (rgthree) enable ControlNet Group + LLM PROMPTER.
Pick the LLM preset: ControlNet_Edit_Mode or K1_Krea2_HighDetailed.
Everything else — as in standard generation.
Panel controls: strength Lora — structure hold (default 0.90; lower = looser, higher = stricter) · resolution limit — max side (default 1920).
Depth is the released control type (canny / pose / normal need a matching Control-LoRA).
🎨 COLORGRADE (LUT) — final look & grain
Grades the final image and gives it that amateur-photo feel.
LUT filters — apply .cube LUT files. Drop your own into ComfyUI/custom_nodes/ComfyUI-OlmLUT/luts/ and they appear in the list.
Tone controls — temperature, hue, brightness, contrast and more, for fine final-look tuning.
Digital grain — grain amount, grain size, color amount — adds realistic sensor noise so renders read as phone photos, not CG.
Standalone editor — you can grade any image from your PC: switch the slider to 2, upload it via upload image for edit, and disable the SAMPLING module in Fast Groups Bypasser.
A|B compare — set the selector to 1 or 2 depending on which photo you're processing, so the before/after view shows correctly.
👉 Model: Selfora — Krea 2 Realistic Amateur Checkpoint
Required custom nodes
See the "Model Links" note inside the workflow — everything is listed with direct download links. Install missing ones via ComfyUI-Manager.
Support
Solo project — the free stuff stays free ☕ Want new models early + a Discord role? Patreon's open, totally optional → https://www.patreon.com/c/ArtfatAI
Credits
Built on moody's Krea 2 base. Thanks to catlover1937 (moody).
Description
AIO — the full Selfism pipeline in one workflow: preset generation + Krea 2 Identity Edit (restage / outfit transfer / masked edits) + Depth ControlNet + Colorgrade (LUT). Preset-driven, no trigger words.
FAQ
Comments (5)
Some feedback. I try to use a character lora, maybe it works better with image input, I don't know. What I noticed, it doesn't matter which preset is used, the LLM output is always the same. Word for word.
Found it — your seed is set to "fixed" (control after generate). That locks the output, so you get the exact same text every run no matter the preset. Switch it to "random" and each preset will do its own thing. 👍
@Artfat It's on randomize in your workflow. I just flipped the switch "LLM enabled" to true. I already reported an issue on your llm git repo, I got it installed by using the cpp wheel. Maybe it's not working properly, but I do not see any hint in the output. The only messages I get are the following (whether I enable load image 1/2 or not):
[ERROR] Failed to validate prompt for output 1023:
[ERROR] * (prompt):
[ERROR] - Required input is missing: images
[ERROR] * SaveImageKJ 1023:
[ERROR] - Required input is missing: images
[ERROR] Output will be ignored
EDIT: It would have worked but your node did not threw an error. Qwen3.5 is not available so the select is None. For a newer version for llama users need to compile it. I build a script for that, but honestly this is too much overhead for the average joe on AMD. Maybe it is easier to mark this workflow in its current form as NVIDIA only. It was necessary to adjust the code of the prompter anyways.
If you are interested I could send you the adjusted repo (I did not publish it, only locally available). But I do not think anyone would take the steps necessary. But with it, your node works.
@daze45230 Really appreciate the detailed report, this is super helpful. And it actually explains your earlier one too. When Qwen3.5 isn't available the node just returns nothing instead of erroring, so the LLM never actually runs. That's why the prompt came out the same every time no matter the preset. The presets were fine, the model was just never loading. My bad on the silent fail.
So that's the first thing I'll fix: the node will throw a clear error when the model isn't loaded, instead of passing nothing down and giving you that confusing "missing images" further along. I'll also switch the default to a model that runs on the shipped llama.cpp, so people don't have to compile a newer build just to get it going. That should cover most AMD folks without the extra steps.
And yes, please send the adjusted repo. Even if I don't merge it as-is, seeing exactly what you changed will save me a lot of time getting AMD working right. Thanks for actually digging into this instead of just dropping it.