MOSS De-Render // Photoreal Negative Slider

Say goodbye to the plastic, airbrushed "3D Video Game" look on MOSS-Image.

When using 4-step or 8-step SFT Turbo LoRAs on MOSS-Image, step distillation tends to wipe out micro-textures. Skin turns waxy and poreless, lighting looks like Unreal Engine 5 ray-tracing, and human subjects end up looking like smooth CGI dolls.

MOSS De-Render fixes this at the root. Trained as an inverse delta patch directly against raw photography with the Turbo adapter active as a Context LoRA, it mathematically subtracts the plastic SFT render traits out of the DiT latent space.

⚡ IMPORTANT: THE 2-LORA DUO SYSTEM (READ FIRST)

In my testing, pushing this De-Render LoRA into higher negative values successfully destroys the 3D plastic look, but it tends to make the image softer and lose sharpness.

To fix this, I created a second companion LoRA: MOSS SharpDetail // Clarity Companion.

They are designed to be run together as a tandem:

• Main LoRA (MOSS De-Render): Run at a NEGATIVE value (e.g. -0.8 to -1.2) to subtract the waxy 3D look.

• Companion LoRA (MOSS SharpDetail): Run at a POSITIVE value (e.g. +0.6 to +1.0) to pull edge definition, sharpness, and clarity back into the photo.

An Honest Note: This tandem setup is far from perfect. Diffusion math has many variables, and you will need to fiddle with both dials. However, in my extensive testing, downloading and stacking both LoRAs together gave the absolute best results. You can certainly use just the main De-Render LoRA on its own if you prefer a softer, candid look, but having both gives you total control over the balance between realism and sharpness.

⚠️ BALANCE IS EVERYTHING (QUALITY DROPS IF OFF)

Image quality WILL degrade if your slider values are out of balance.

Because these models act as push-and-pull vectors, there is a sweet spot you must dial in:

• If you don't push De-Render negative enough (e.g. -0.3): The Turbo engine wins, leaving the image smooth, waxy, and synthetic.

• If you push De-Render too negative without enough sharpness compensation: The image can lose focal clarity or become overly soft and noisy.

• When stacking additional character/physique LoRAs: Every extra LoRA shifts this balance point.

None of these numbers are written in stone. Treat the numbers below as starting baselines to fiddle from.

🔑 Trigger Words

The sliders do the mathematical work through their weight values, but adding these trigger words into your positive prompt steers the restored details into authentic camera realism:

code Text

real photograph, candid shot, authentic skin texture, natural ambient lighting, raw unedited photo

What each keyword targets:

• authentic skin texture: Breaks up poreless plastic skin and restores natural pores, sweat sheen, and micro-geometry.

• raw unedited photo / real photograph: Suppresses digital over-sharpening and CGI post-processing.

• natural ambient lighting: Replaces artificial ray-traced bloom and plastic highlights with authentic room or sun light.

• candid shot: Relaxes stiff, mannequin-like poses into natural posture.

⚡ Recommended LoRA Stacking Order (In ComfyUI)

MOSS-Image’s 5B DiT handles multi-LoRA stacking exceptionally well if you keep text conditioning untouched.

The Full Stack:

1. Speed Layer: MOSS Turbo SFT LoRA (at +1.0) for fast 4 to 8 step generation.

2. De-Render Layer: MOSS De-Render Slider (at -0.8 to -1.2) to strip away plastic smoothing.

3. Clarity Layer: MOSS SharpDetail Companion (at +0.6 to +1.0) to recover edge definition and crispness.

4. Subject Layer (Optional): Your Character, Muscle, or Outfit LoRA (at standard positive weights +0.7 to +1.0).

Always use LoraLoaderModelOnly nodes in ComfyUI so you only modify the DiT transformer blocks, keeping Qwen3.5 text conditioning clean.

🎚️ Weight Starting Baselines (Fiddle from here)

Main De-Render LoRA (Always Negative):

• -0.4 to -0.7 (Light Touch): Softens digital highlights and smooth gradients. Great for simple portraits where you don't need heavy compensation.

• -0.8 to -1.1 (Standard Sweet Spot): The go-to baseline to cancel the Turbo plastic look while retaining natural skin micro-contrast.

• -1.2 to -1.5 (Heavy Compensation): Use when stacking multiple heavy LoRAs that introduce their own plastic smoothing.

SharpDetail Companion LoRA (Always Positive):

• +0.4 to +0.6: Gentle sharpness recovery.

• +0.7 to +1.0: Standard pairing with De-Render at -1.0. Brings back crisp eyes, eyelashes, and clothing fibers.

• +1.1 to +1.3: For wide shots or high-resolution portraits where you need maximum edge definition.

⚙️ Recommended Generation Settings (MOSS + Turbo)

• Steps: 6 to 8 steps (4 steps works at native 1024x1024, but 6 to 8 steps provides the sampling headroom needed for stacked DiT layers to resolve fine textures at higher resolutions).

• CFG: 1.0 (Turbo distillation requires CFG 1.0; going higher will destroy the image).

• Sampler: euler

• Scheduler: simple (or sgm_uniform)

• Latent Node: EmptyFlux2LatentImage (16-channel)

• VAE: flux2-vae.safetensors

💡 Can I use them WITHOUT the Turbo LoRA?

Yes. You can run the De-Render slider at -0.4 to -0.7 on standard base MOSS-Image (20 to 30 steps) as a general de-beautifier to pull the model away from digital art and toward candid camera photography. If you find the output slightly soft, dial the SharpDetail companion in at +0.4 to +0.6.

💬 Summary & Community Feedback

Start with De-Render at -0.9 and SharpDetail at +0.7 with the trigger words in your prompt. If it looks too soft, push SharpDetail higher or ease De-Render back toward -0.7. If it still looks like an Unreal Engine render, push De-Render down to -1.2.

This is an experimental tandem patch—I might be wrong on some combinations, so feel free to experiment and find what works best for your specific workflow.

Please share your workflows, stacked chains, and side-by-side comparison images in the gallery below!