I haven't tested 2509, but I suspect it will just with diminishing returns. It does work pretty well with 2511 edit.

Dataset was 260 images, I did a low rank for a slow stable burn. Still refining my understanding of this, what would you suggest for 260 images? I also did batches of 12 images for stability.

1️⃣ The lighting LoRA is overriding detail‑preserving layers

What’s happening:
Many lighting‑or‑color‑grade LoRAs are trained to modify the overall tonal balance of an image. They often do this by pushing the model toward a smoother, more “glow‑y” look. When you stack that on top of a detail‑heavy LoRA (like the Qwen‑2512 General Nudes & Sex Stuff), the smoothing effect can bleed into edges and fine textures, making the final picture look soft or out‑of‑focus.

What to try:

Reduce the strength (alpha) of the lighting LoRA. Start with a value around 0.2–0.3 and increase only if you need more light effect.

Apply the lighting LoRA after the main generation step (e.g., use it as an “up‑scale/finetune” pass rather than part of the initial prompt). In many pipelines you can run a second pass with just the lighting LoRA at a low strength.

2️⃣ Sampling steps / CFG scale mismatch

What’s happening:
If you’re using a relatively low number of diffusion steps (e.g., 20–30) or a very high CFG (classifier‑free guidance) value (>12), the model can converge too quickly and produce a “soft” result, especially when multiple LoRAs are fighting for control.

What to try:

Increase sampling steps to 50–80 (or even 100 if you have GPU headroom). More steps give the network more opportunity to refine edges.

Lower CFG to somewhere between 7–9. This lets the LoRAs express their learned details rather than being forced into a single, overly‑sharp interpretation.

3️⃣ Resolution & Latent Upscaling

What’s happening:
If you generate at a low base resolution (e.g., 512×512) and rely on a post‑upscale (ESRGAN, SwinIR, etc.) that isn’t tuned for the specific LoRA, the upscaler can introduce softness.

What to try:

Generate at a higher native resolution (768×768, 1024×1024, or whatever your GPU can handle) before applying any external upscaler.

If you must upscale, use a high‑quality upscaler with a “sharp” preset (e.g., “Real‑ESRGAN x4 + sharp”).

4️⃣ Prompt wording & Negative prompts

What’s happening:
Sometimes the prompt itself asks for “soft lighting”, “dreamy”, or “glowy” attributes, which reinforce the smoothing effect of the lighting LoRA.

What to try:

Add a negative prompt such as blur, out of focus, soft, dreamy to explicitly discourage softness.

Include positive terms that promote crispness: sharp details, high definition, 8k, fine texture.

5️⃣ Model & LoRA compatibility

What’s happening:
The Qwen‑2512 General Nudes & Sex Stuff LoRA was trained on a specific base model (e.g., SDXL‑1.0 or SD‑2.1). If you’re using a different version (or a custom checkpoint), the weight alignment can be off, resulting in artefacts like blur.

What to try:

Confirm you’re using the exact base model mentioned on the Civitai page.

If you must use a different checkpoint, try merging the LoRA with a lower strength (e.g., 0.5) and see if clarity improves.

QUICK “cheat‑sheet” checklist

StepActionTypical Value
1️⃣Reduce lighting LoRA strength0.2–0.3
2️⃣Increase diffusion steps50–80
3️⃣Lower CFG scale7–9
4️⃣Raise base resolution768×768+
5️⃣Add negative prompts for blurblur, out of focus
6️⃣Verify base model versionMatch Civitai page
7️⃣Use a sharp upscaler (if needed)Real‑ESRGAN + sharp

Example prompt & settings

Run a single pass with those values, then optionally run a second “refine” pass using only the lighting LoRA at strength 0.15 if you still want a touch more glow.

TL;DR

Blur usually comes from an overly‑strong lighting LoRA, low steps/high CFG, low resolution, or mismatched base model.

Dial down the lighting LoRA, bump up steps, lower CFG, generate at higher native resolution, and add a negative “blur” prompt.

Check that you’re on the correct base checkpoint; otherwise reduce LoRA strength.

Give those tweaks a try and let me know how it turns out! If you still see issues, feel free to share your exact settings (steps, CFG, strengths, resolution) and I can help fine‑tune further. Happy generating!

Good note, thank you! Yeah, I need to do a bit more post processing to my images. I think the next run I do I will process all of them through Qwen Edit 2511 to remove things I don't want like watermarks. I'm able to get decent results with the lightning loras, but I'll five those a shot.

@TimmyHodor I did a multi-pass method. First process them all with
qwen3-vl-8b-nsfw-caption-v4.5, then manually clean up anything the model did wrong as well as add common keywords (ex. reverse cowgirl) that the LLMs just don't get.

@ChillDesire Generally I drop the CFG a lot with the lighting LoRa's when I'm seeing the high contrast / body horror stuff. I noticed that Qwen does pretty well all the way down to 0.6 CFG and still adheres to the prompt pretty well assuming that it's well written. I've also noticed you can get great results going higher than the LoRa steps prescribe. So use 8 steps on the 4 step LoRa.