FameGrid for Z-Image & Qwen-Image

Fine-tuned on a meticulously curated dataset of modern social-media influencer photography. This LoRA wasn’t slapped together — it took weeks of curation, iteration, and a painful amount of server burn to get natural skin, modern lighting, and clean composition to land consistently.

New UltraMuse Discord

We’re building the most vibrant creator hub around — workflows, free resources, tutorials, and experiments drop there.

👉 Join UltraMuse Discord (For Official Workflows, Support, Exclusive releases)
---------------

🧪 Z-Image 0.5 Beta (Experimental)

A quick note on the Z-Image 0.5 beta release.

This version is labeled beta for a reason.

Known Limitations

• Anatomy issues, most noticeable in feet

• Text rendering is significantly weaker than the base Z-Image model

These are known problems and are expected to be addressed in the next release.

Why Release 0.5?

This build introduces a strong stylistic direction and is intended for:

• Aesthetic exploration

• Prompt and workflow testing

• Early access for creators experimenting ahead of full refinement

If you need clean anatomy or reliable text, wait.
If you want to explore where Z-Image is heading, this is it.

👉 Grab the 0.5 workflow in the UltraMuse Discord

🆕 1.5 Latest Updates (Qwen)

New Trigger Word: rlskn

Use this trigger along with the default one "IGMODEL" to unlock more realistic skin textures, natural imperfections, and average-looking subjects. Perfect for authentic lifestyle content or grounded social media aesthetics.

Important Notes

[Download 1.5 Workflow Here (Qwen)]

Start your prompts with something like:
“iPhone photo of an IGMODEL woman with rlskn (detailed skin:1.3) …”

Additional guidance:

• Use rlskn for realism, texture, and believable skin

• Lower CFG = more realism

• Higher CFG = more sharpness, but can reduce naturalness

• Match CFG and LoRA weight:

  • Low CFG → low weights

  • High CFG → higher weights

• This model loves long, descriptive prompts — 800+ characters is ideal

• If your outputs aren’t realistic enough, boost LoRA strength

Who Is This For?

FameGrid was trained on imagery emphasizing:

• Natural, confident posing

• Clean, modern composition

• Soft but controlled lighting

• Professional yet relatable social media aesthetics

Ideal for:

• Lifestyle & influencer-style photography

• Product shots with human context

• Virtual character creation

• Social media content generation at scale

Trigger Words

• IGMODEL — The main trigger for influencer-style realism

• rlskn — Enhanced skin texture and natural appearance

Recommended Settings (Qwen-Image)

• Model Flow: 2

• Sampler: res 2s

• Scheduler: simple

• CFG: 1–3 (1 recommended)

• LoRA Weight: 1.0–2.0 (1 recommended; increase if realism drops)

FameGrid Qwen Checkpoint (Coming Soon)
Full checkpoint will be released on CivitAI once training completes.

🧪 Behind the Scenes — The Training Gauntlet

Here’s the part most creators won’t tell you: FameGrid didn’t come together cleanly. It was a brutal training cycle — models collapsing, overfitting, underfitting, and entire runs being trashed because one stylistic artifact ruined the whole aesthetic.

I destroyed more models than I kept. I restarted pipelines over and over. Re-curated images, rebuilt datasets, reran experiments that burned through cloud credits like they were free. It was expensive and punishing — but that’s the price of chasing realism without slipping into uncanny, plastic, or over-processed territory.

Every iteration forced a decision: throw it out or settle?
I never settled.

The final result is a LoRA that feels natural, modern, and clean — the kind of influencer photography that actually passes in real social-media environments.

Was it worth it?
Yeah. Absolutely.

🛠 Additional Notes & Upcoming Tools

FameGrid is highly sensitive to prompt structure, phrasing, and stylistic cues. Tiny prompt differences can create massive outcome changes — that’s part of why the realism is so strong, but it also means casual prompting won’t unlock its full potential.

To solve that, we built:

• FameGrid Prompt Crafter

A tool that automatically generates optimized prompts tailored to the model’s strengths.

• FameGrid Prompt Expander

Takes short prompts and expands them into long-form descriptions that match the model’s ideal prompt length and composition.

Both will be released in the UltraMuse Discord within the next few days.