Dissolve image lighting Kontext V2

What this workflow is: A Flux Kontext-powered product relighting and scene-blending pipeline. It takes a product image, reconstructs its lighting to match a target scene with luxury advertising quality, restores fine material detail via frequency separation, and optionally upscales to commercial resolution.

Two groups organize the canvas: "Click to Run" handles the core generation, and "HD Upscale" handles the optional hi-res refinement pass.

What this workflow does (at a glance)

It takes a product image, runs it through Flux Kontext with a custom relighting LoRA, reconstructs sophisticated studio-quality lighting on the product surface, then restores high-frequency material detail using frequency separation — outputting a clean, upscale-ready result suitable for e-commerce or luxury advertising.

What the node chain is doing (in practical terms)

1) Input & preprocessing

• LoadImage feeds the product photo

• ConstrainImage caps it at 2048×2048 before encoding to avoid VRAM blowouts

2) Core model stack (Flux Kontext)

• UNETLoader loads flux1-kontext-dev — the edit-capable Flux model

• DualCLIPLoader loads clip_l + t5xxl_fp8 for dual-encoder text conditioning

• VAELoader loads the standard Flux AE (ae.sft)

• LoraLoaderModelOnly injects the Adaptive Dissolve/Blend LoRA (自适应溶图V3, strength 1.1) — the key driver of lighting reconstruction behavior

• FluxGuidance sets guidance to 3.5

3) Relighting prompt

• CLIPTextEncode encodes a carefully engineered prompt that instructs Flux to:

  • Reconstruct directional hard light with luxury-advertising character on key product surfaces

  • Maintain a soft ambient base layer for e-commerce brightness standards

  • Add sharp highlights, micro-reflective surface detail, and natural contact shadows

  • Generate controlled reflections on smooth/glass/metal surfaces, matched to the main light direction

  • Strictly preserve the product's original color, hue, proportions, and geometry — no color drift allowed

4) Reference-conditioned generation

• ReferenceLatent feeds the encoded input as a latent reference into the conditioning chain, anchoring the edit to the product's original structure

• KSampler runs 30 steps at denoise 1.0 (full generation from the reference), producing the relit image

5) Detail restoration — the frequency separation pass

• RMBG (RMBG-2.0) generates an alpha mask isolating the product from its background

• LayerUtility: ImageMaskScaleAs rescales the generated result to match original input dimensions

• LayerUtility: HLFrequencyDetailRestore composites high-frequency detail (texture, fine edges, material grain) from the original image back onto the relit result, masked to the product region

• This is the critical step that prevents the generation from losing fine surface texture — fabric weave, metal grain, label print, etc. all survive

6) Output & comparison

• SaveImage saves the final composited result

• Image Comparer (rgthree) shows a live before/after slide comparison directly in the canvas

7) Optional HD upscale pass (disabled by default)

• UpscaleModelLoader → ImageUpscaleWithModel runs a model-based 2× upscale using 2xNomosUni_span_multijpg_ldl

• ImageScaleToTotalPixels normalizes the result to 8 megapixels via Lanczos

• TTP_Image_Tile_Batch + TTP_Tile_image_size split the upscaled image into overlapping tiles (2×3 grid, 5% overlap) for tile-level refinement

• ImpactImageBatchToImageList / ImageListToImageBatch handle list↔batch conversion around the tile processing loop

• A second KSampler (6 steps, denoise 0.4, Euler/kl_optimal) runs per-tile refinement using a separate e-commerce 3D rendering LoRA (strength ~0.39)

• VAEDecodeTiled decodes with 1536px tiles to stay within memory limits

• Tile stitching uses stored position and grid metadata from TTP_Image_Tile_Batch to seamlessly reassemble the final image

The "big win": geometry-locked, detail-preserving relighting

Most product relighting workflows fail in two places: the generation drifts from the original product geometry, or it smears fine surface detail during synthesis. This workflow solves both by:

• Using Flux Kontext with a latent reference to lock in the product's original shape and proportions

• Running a frequency separation restoration step that injects original high-frequency texture back onto the relit surface — so the lighting changes, but the material stays sharp

The result meets e-commerce clarity standards while carrying the contrast and depth of luxury advertising photography.

Where this shines (real use cases)

• E-commerce product shots: Drop any product in, get back consistent studio lighting with clean shadows and sharp material texture

• Scene matching: Relight a product to match the ambient character of a target environment without reshooting

• Batch catalog work: Tune the LoRA strength and prompt once, run multiple products through the same lighting treatment for a consistent collection look

• Hi-res output: Enable the upscale branch for print-ready or large-format deliverables without losing fine detail