V4.3_EXP released...
This is based on realism and currently is under testing... More images will be uploaded later... V5 will come soon for more underrepresented concepts..
Strength: 0.8 -> 1 -> 1.2 -> 2


It's my first trained LORA based on KREA 2
V4.1 RELEASED!!
Better prompt adherance. FAR FAR BETTER realistic skins. Face expressions are here. Most things are better..
Things bad:
- Masturbation concept not defined completely

i tested it with filterbypassv3 and projector scale loras without any issues but they're completely unnecessary now compared to KNP v2

@19doorsside884 I just used it in combination with a self-trained character lora. from 0.6-1 character bleed starts and gets worse the higher. Its not crazy but there.

I tried V4 with some of Turbo's built-in names and just V4. The bleeding is definitely there but I am really anal about likeness so it may not bother other people. I didn't see bleeding in the earlier versions.

@19doorsside884 Another detail that i confirm is that it tends to blur everything that isn't in the foreground. In a simple image where a girl with her legs spread is very close to the camera, her pelvis is sharp, but her torso, face, and the rest of the background are blurred, or in other words, have a bokeh effect.

basically, save one of these image on your desktop, then drag and drop the image into your comfyui, correct the models on the red node errors with your correct path, eventually go on manager and update missing custom nodes, and your go to go. Yes Krea 2 working on image2image, but you have to fine a specific already made worflow, no idea if this is present on Comfyui template, and porbably if is there is SFW, only, so do a bit of reserach here on Citivai type krea 2, then select on your research "worflow" you should see many of them. :)

Try creating a new folder and cursor project and paste this in. It's a one shot prompt my agent wrote after setting it up for me. I used windows and a 5090. ~60s for 1024x1024 generations. I used cursor's default composer 2.5 fast model.

# One-shot setup prompt — Krea2NsfwV4
Copy everything below the line and give it to another person, or paste it into an AI coding agent on **their** machine. Adapt paths and OS as needed.
---
## Prompt (copy from here)

Set up Krea2NsfwV4 locally: a Gradio text-to-image app that runs the Krea 2 NSFW V4 LoRA KNP v4.3_EXP) on top of Krea 2 Turbo using diffusers Krea2Pipeline). Goal: load the LoRA, auto-download the base model on first run, generate images from a browser UI, log load/generate timings, and optionally offload VRAM after each run.

Work on whatever OS, Python env, and GPU the user already has (Windows/macOS/Linux; conda or venv). Do not assume ComfyUI unless diffusers cannot run — prefer the diffusers path below.

### 1. Hardware check (do this first)

- GPU: NVIDIA with CUDA strongly recommended. Note VRAM (GB) and whether anything else is using the GPU.

- RAM: 32 GB+ recommended (base stack is ~33 GB on disk; offload uses system RAM).

- Disk: ~40 GB free for first-run Hugging Face download + LoRA + outputs.

- If VRAM ≤ 32 GB: you must use sequential CPU offload (not whole-model CPU offload). Full bf16 stack is ~33 GB and will not fit entirely on a 32 GB card.

- If VRAM ≥ 48 GB: try full GPU load first; fall back to sequential offload on OOM.

- If no CUDA: warn the user; CPU-only will be impractically slow.

Record GPU/RAM/driver in hardware_specs.txt in the project folder.

### 2. Get the LoRA (manual download — not on Hugging Face)

- Model page: https://civitai.red/models/2725430/krea-2-nsfw-v4

- Version: v4.3_EXP

- File: KNP_000003000.safetensors (~436 MB)

- Mirror API: https://civitai.com/api/v1/models/2725430

Place the file at:

```

<Krea2NsfwV4>/models/loras/KNP_000003000.safetensors

```

Create model_link.txt with one line: https://civitai.red/models/2725430/krea-2-nsfw-v4

### 3. Base model (auto-download on first load)

- Hugging Face repo: https://huggingface.co/krea/Krea-2-Turbo

- Loaded in code as: krea/Krea-2-Turbo via Krea2Pipeline.from_pretrained(...)

- Size: ~33 GB (transformer ~24 GB + Qwen3-VL text encoder ~8 GB + VAE ~0.5 GB)

- Cache location (default): ~/.cache/huggingface/hub/models--krea--Krea-2-Turbo/ (or %USERPROFILE%\.cache\huggingface\hub\... on Windows)

User does not need to download this manually unless they want offline/air-gapped setup (then mirror that HF repo into the cache or a local folder).

### 4. Python environment

Use the user’s existing conda env if it already has PyTorch + CUDA, or create a venv in the project.

Minimum packages:

```

torch>=2.5.0 # CUDA build matching their driver

diffusers>=0.39.0 # must include Krea2Pipeline

transformers>=4.50.0 # Qwen3VLModel

accelerate>=1.0.0

peft>=0.18.0

safetensors>=0.4.0

pillow>=10.0.0

gradio>=5.0.0 # 5.x: pass theme/css to gr.Blocks(); 6.x: check migration guide

```

Verify before building UI:

```python

from diffusers import Krea2Pipeline

from transformers import Qwen3VLModel

import gradio as gr

print("OK", gr.__version__)

```

### 5. Project layout to create (or copy from reference repo)

```

Krea2NsfwV4/

app.py # Gradio UI and launcher

model_loader.py # Krea2Pipeline + LoRA load/generate + timing helpers

models/loras/ # KNP_000003000.safetensors

outputs/ # generated PNGs

logs/ # run_YYYYMMDD.log (timings; prompts NOT logged by default)

tests/test_generate.py # smoke test

run.bat / run.sh # start UI

stop.bat / stop.sh # stop UI / free GPU

requirements.txt

model_info.txt # Civitai digest + prompt guide

example_prompts.txt # copy-paste prompts for this LoRA

README.md

```

Core inference settings (defaults):

| Setting | Value |

|---------|-------|

| Base | krea/Krea-2-Turbo |

| LoRA adapter name | knp |

| LoRA strength | 1.0 (1.0–1.5 for stronger) |

| Steps | 8 |

| CFG | 0.0 |

| Resolution | 1024×1024 |

| dtype | bfloat16 |

| Memory | enable_sequential_cpu_offload() if VRAM ≤ ~32 GB |

*model_loader.py must:**

- Find LoRA in models/loras/

- Load pipeline + load_lora_weights + set_adapters

- Expose format_duration() / timing fields for logs

- Return generation stats (duration, per-step, GPU before/after)

- Use sequential CPU offload on tight VRAM (never whole-model offload on 32 GB — it is ~7× slower)

UI must:

- Run on 127.0.0.1:7862 (change port if taken)

- Buttons: Load/Reload, Offload, Generate

- Checkbox: Keep model loaded (unchecked = unload after each gen to free VRAM)

- Show load time, last gen time, per-step time in status/logs

- Do not send prompts to any remote service; do not log prompt text unless user explicitly asks

Gradio version note: On Gradio 5.x, put theme= and css= on gr.Blocks(...), not launch(). On 6.x, check the migration guide.

### 6. Run sequence

```bash

cd Krea2NsfwV4

# optional: python -m venv .venv && source .venv/bin/activate (or .venv\Scripts\activate)

pip install -r requirements.txt # only if deps missing

python tests/test_generate.py # first run downloads ~33 GB; then one test image

python app.py # or run.bat / ./run.sh

```

Browser: http://127.0.0.1:7862 → Load / Reload model → paste prompt → Generate.

Example prompt (from this LoRA’s guide):

```

wide side profile shot, full body visible, nude adult woman standing in a messy bedroom, cheap smartphone camera, harsh direct flash, very realistic

```

More examples: see example_prompts.txt in the reference project.

### 7. Prompt rules (important for this LoRA)

- Short, punchy sentences — not long tag lists or “8k masterpiece cinematic”

- Formula: [camera angle] + [subject/action] + [detail] + [cheap camera / harsh flash] + "very realistic"

- Prefer geometric full-body angles (wide side profile, 45° front, overhead) — avoid POV

- Multiple people: “The first is… The second is…”

### 8. Done when

- [ ] LoRA file in models/loras/

- [ ] Smoke test writes PNG to outputs/ and prints load + generate duration

- [ ] UI loads without Gradio theme/css errors

- [ ] One successful generation from the UI

- [ ] logs/run_*.log records load_duration and generate_duration

- [ ] User told where base model caches and that first download is ~33 GB

### 9. Optional alternatives (only if diffusers path fails)

- ComfyUI + FP8 weights: https://huggingface.co/Comfy-Org/Krea-2 — often faster on 32 GB cards but different workflow.

- Do not install extra packages blindly; check conda/ComfyUI first per user preference.

Report back: GPU model, VRAM, memory mode used, first-load time, generate time (8 steps @ 1024²), and paths to LoRA + HF cache.

---

## End of prompt

Reference implementation (if the user has access): opensource-img-gen/Krea2NsfwV4/ in the sandbox repo.

A few things to try:
- Krea2 likes to generate at 1920x1088 or larger; so any HD resolution where each side is divisible by 16 (yes, 16, not just 8) can help. This is because Krea2 uses a Qwen VL text encoder which natively splits all latents into 16 channels for processing.
- Speaking of Qwen technology; that crosshatch or "deinterlacing" pattern you see over the images is a by product of the default VAE. I got around this by using https://civitai.red/models/2739236/qwenimagevaeliquid1127 and the difference was astounding. (Just remember to use the default VAE for any image Encoding [img2img] because of that whole 16 channels thing I mentioned) Liquid1127 is just for the final Decoding step.
- Experiment with different samplers and schedulers. RES4LYF should be a required install for all ComfyUI users because it unlocks the bong_tangent scheduler ( it uses a tangent function to control the transition which allows the AI model to quickly define the main composition early on and spend the remaining steps refining fine details). Also try samplers than end with "cfg_pp" They're meant for lower guidance scales (like a 1.0 CFG Turbo model) while maintaining exceptional prompt adherence, producing less saturated/burnt colors, and drastically improving image invertibility.
- Make sure your workflow is using nodes that pass CLIP with the LoRAs and not just Model Only

Yeah, mystic xxx more plastic, this one more realistic but imo sometimes ignores half the prompt to just go and add dicks everywhere and when you're asking just for a girl and a man naked but not doing anything particular she will be sucking his dick or getting fucked, etc

Still, amazing work !

@19doorsside884 ОК, Agreed) And please don't forget about sperm, there's a damn shortage of it.

bonus!

I even got this when I created my own lora. I think there was one or maybe two bath/wet skin pictures in the dataset, but it still causes many of my generations to come out with weird wet skin or large droplets on the skin. IDK if it's in the training or something inherent in Krea2.

The way I get around it is to use loraholic's wetness slider lora and set the strength to -3.

@DaveNotDavid I have this with other lora's as well, but not always. So i'm not sure if there's a trigger for it or if its something else, but the wetness slider does help.

@az420 not exactly, i switched over to mystic xxx getting better results with that.