BXR - T2I Upscaler & Color Grade (V3 Refresh) 👑

Welcome to the official V3 refresh of the BXR - T2I Upscaler & Color Grade workflow! This is the core foundation I have used from day one, now polished and updated with some slight changes to make it even more robust, efficient, and hardware-friendly.

This setup is engineered to generate highly detailed, crisp imagery from scratch, utilizing an advanced multi-stage upscaling pipeline and precision color grading to give your final renders a professional, cinematic pop.

🚀 What's New in V3

• Integrated VRAM Purging: Added Easy Clean VRAM Used nodes right after the VAE decode to clear GPU memory before the heavy upscale, preventing OOM crashes on standard consumer hardware.

• Refined Ultimate SD Upscale Pipeline: Calibrated to a 3.0x scale using 4x_NMKD-Siax_200k.pth with Chess tiling (1536x1536) and a 48-pixel mask blur for seamless, artifact-free high-res outputs.

• Hardware-Accelerated Color Grading: Swapped prompt-based contrast tweaking for the FL Image Adjuster node, locking in rich saturation and contrast (~10.18) for cinematic vibrance.

• Optimized KSampler Routing: Split prompting into base generation and refiner conditioning to isolate structural rendering from texture enhancement.

⚙️ Workflow Breakdown

The core philosophy of this workflow is balancing high-fidelity generation with optimized VRAM management.

[Base Generation] ➔ [VAE Decode / Preview] ➔ [VRAM Purge] ➔ [Ultimate SD Upscale] ➔ [Color Grading Layer]

1. Resource Loading & LoRA Management

• Load Checkpoint: Built to leverage high-tier realism models (specifically optimized for checkpoints like illustriousRealismBy_v10VAE.safetensors).

• Power LoRA Loader (rgthree): Stacks multiple character, utility, and detail-enhancing LoRAs cleanly without cluttering the workspace. Balances structural enhancements (like DetailerILv2) alongside concept-specific weights.

2. Latent & Initial KSampler

• Empty Latent Image: Starts at a high-resolution base aspect ratio of 1200 x 1600 with a batch size of 3.

• KSampler (Low Res Preview): Runs for 25 steps at a CFG of 6.0. It uses the dpmpp_2m sampler paired with the karras scheduler to lock in the composition and core details efficiently before upscaling.

3. Ultimate SD Upscale Pipeline

The heavy lifting is handled by the Ultimate SD Upscale node, scaling the image up by a massive factor of 3.0x.

• Denoise Control: Kept low at 0.35 to ensure the upscale strictly enhances details, skin texture, and sharpness without shifting or morphing the original composition.

4. Post-Processing & Color Grading

• FL Image Adjuster: Mathematically boosts Saturation and Contrast to eliminate the typical washed-out neural output look, delivering deep, vibrant tones right out of the generator.

📊 Recommended Settings Matrix

Base Generation Stage

• Steps: 25

• CFG: 6.0

• Sampler: dpmpp_2m

• Scheduler: karras

• Denoise: 1.00

• Target Dimensions: 1200 x 1600

Ultimate SD Upscale Stage

• Steps: 26

• CFG: 7.0

• Sampler: dpmpp_2m

• Scheduler: karras

• Denoise: 0.35

• Target Dimensions: 3.0x Scale (~3600 x 4800)

📝 Prompting Strategy & Guidance

This workflow splits prompting into primary generation blocks and specific refiner blocks to maximize control:

• Primary Prompt: Focus heavily on your core subject matter, clothing, environmental background, specific physical attributes (e.g., maintaining a "Mild" toned level for a character's physique rather than overblown muscle), and lighting moods.

• Refiner Positive Prompt: Do not paste your entire descriptive prompt here! Keep it strictly focused on quality-boosting modifiers, hyper-detailed skin textures, photographic clarity, and eliminating artifacts (e.g., score_9, score_8_up, masterpiece, photorealistic, high-fidelity details).

• Negative Prompts: Standard embedding blocks filter out anatomical mutations, compression artifacts, and poor framing.

🛠️ Required Custom Nodes

Before running, ensure you have the following custom node suites installed via your ComfyUI Manager:

• rgthree-comfy (Power LoRA Loader)

• UltimateSDUpscale

• ComfyUI-Easy-Use (VRAM purging utilities)

• Fill-Nodes (FL Image Adjuster)

Note: Verify your models (illustriousRealismBy_v10VAE.safetensors and 4x_NMKD-Siax_200k.pth) are placed in their respective checkpoints and upscale_models directories before queuing your first prompt!