User Guide: ULTRA_REALISM_4STEP_LCM_SIMPLE
by Vivek Chauhan (vivekchahaman600 / ChauhanVivek)

Overview

ULTRA_REALISM_4STEP_LCM_SIMPLE is an SDXL 1.0 full checkpoint distilled for ultra-detailed photorealism in just 4 steps.

It is a complete checkpoint (not a LoRA). The LCM (Latent Consistency Model) distillation and the “simple” scheduler are baked into the weights. Speed and the photorealistic look are already built in — you only get the intended quality when you match the exact schedule it was trained/distilled for.

Best for: Portraits, full-body photography, street and interior scenes, editorial stills, and character work that should read as camera-captured, with visible skin texture, natural light, and realistic materials.
Not intended for: Painterly, oil, watercolour, pencil, or anime styles. Use your matching Chauhan style checkpoints or stack a light style LoRA for those.

Critical Settings (Do Not Deviate)

These are required. Using the “normal” SDXL recipe will produce poor results.

SettingValueNotesCheckpointULTRA_REALISM_4STEP_LCM_SIMPLELoad as the model, not as a LoRASamplerLCMRequired. Not Euler, not DPM++ 2MSchedulersimpleRequired. Not Karras, not exponentialSteps4Required. Only trial 5–6 if neededCFG Scale1.0Required. Stay 1.0–1.5. Above ~2 the path breaksDenoise1.00Full generation. Lower only for img2imgClip Skip1Do not leave at 2 (SD 1.5 habit)Resolution1024×1024 or 832×1216 (and similar)Stay near 1 megapixel for best results

ComfyUI KSampler exact values:
steps: 4 · cfg: 1.0 · sampler_name: lcm · scheduler: simple · denoise: 1.00

Why These Settings Matter

• At CFG 1.0, classifier-free guidance is effectively off. The negative prompt does almost nothing. Write the entire picture in the positive prompt.

• If a defect refuses to leave, raise CFG only to 1.2–1.5 and keep the negative short. Never climb to 5–7.

• Longer schedules or different samplers do not “improve” this model — they break the distilled path.

Recommended Workflow

1. Basic Text-to-Image (Forge / A1111 / ComfyUI)

• Load the checkpoint.

• Set sampler = LCM, scheduler = simple, steps = 4, CFG = 1.0.

• Resolution: 1024×1024 or portrait/landscape near 1 MP.

• Positive prompt only (or very light negative).

• Generate.

2. Face Enhancement

• ADetailer (or equivalent face detailer) is highly recommended for solo portraits.

• Keep denoising strength low (under ~0.3–0.4) to avoid face drift.

3. Upscaling / High-Res

• First pass at native resolution (1024-class).

• Then use your preferred upscaler (Ultimate SD Upscaler, 4x-UltraSharp, etc.) with low denoise (0.15–0.35) if refining.

• Because the base is already sharp and realistic, heavy second-pass denoising is usually unnecessary.

4. Img2Img

• Same sampler/scheduler/steps/CFG.

• Lower denoise according to how much you want to change the source (typical 0.3–0.6).

Prompting Tips

• Write clear, photographic descriptions.

• Emphasize lighting, skin texture, camera, and materials:
  photorealistic, realistic skin texture, natural lighting, sharp focus, 35mm, f/1.8, detailed pores, ordinary daylight…

• Keep prompts relatively focused. Overly complex or conflicting prompts can reduce realism.

• Negative prompts are largely ignored at CFG 1.0. Only use a short negative if you raise CFG slightly.

• Works well with character LoRAs (including your own such as Vivekamarisvoss) at moderate strength (0.6–0.9). Prefer one strong identity LoRA at a time.

Common Mistakes to Avoid

• Using DPM++ 2M Karras + 20–30 steps + CFG 5–7 → broken results.

• Leaving Clip Skip at 2.

• Treating it like a regular SDXL or Lightning model.

• Raising steps or CFG “to get more quality.”

• Expecting painterly or anime output.

Compatibility Notes

• Forge / Automatic1111: Full support. Ensure LCM sampler is available (it is in modern builds).

• ComfyUI: Native. Use the standard KSampler with the exact values above.

• VAE: Use the built-in or a high-quality SDXL VAE (auto is usually fine).

• Compatible with most SDXL ControlNets, IP-Adapters, and detailers when used carefully.

• File is a standard safetensors checkpoint (~6.46 GB typical for SDXL).

Quick Reference Card

text

Model:     ULTRA_REALISM_4STEP_LCM_SIMPLE (SDXL)
Sampler:   LCM
Scheduler: simple
Steps:     4
CFG:       1.0 (max 1.5)
Clip Skip: 1
Res:       ~1024×1024

This model is designed for speed and camera-like realism. Match the schedule, write clean photographic prompts, and it delivers sharp, detailed, photorealistic images in four steps.