FLUX Tiled Upscaler Plugin: Solving Transformer Fragment Stitching
Resident_Ad7247 · reddit · 2026-08-08
The author developed a ComfyUI tiled upscaling node for FLUX.2 klein and similar models. Because FLUX uses a Transformer architecture with absolute position embeddings (RoPE), directly applying traditional UNet latent-slicing tricks causes the model to perceive each tile as a small, complete image, redrawing the entire subject at wrong scales inside every tile.
After attempts to patch RoPE shifts failed, the author found success by tiling in pixel space instead of latent space:
- Each tile is cropped from the canvas of already-generated neighbors to ensure continuity of real pixels.
- Per-tile color matching prevents exposure or tint drift.
- Blend weights are calculated based on actual per-side overlap.
This approach handles all continuity and blending outside the model, treating every model call as a complete image at native resolution. The node is available on GitHub and the ComfyUI Registry.
More from Multimodal
- Local Video Generation: MiniMax Lags Far Behind LTX in Inference Speed — PhilosopherSweaty826 · 2026-08-08
- Micro-expressions Matter More Than Pixel Count in AI Video — aftahi_ai · 2026-08-08
- Runway Launches Seedance 2.5: 30-Second Videos with Sound from 50 References — JeffSynthesized · 2026-08-08
- Seedance 2.5 generates realistic UGC videos, accelerating brand ad creative automation — Aiden_Tech_Ai · 2026-08-08
- MiniMax H3 vs. Seedance 2.0: A Side-by-Side Video Generation Test — ajrss2009 · 2026-08-08
- Running Minimax H3 on 12GB VRAM: 9-Min Local Video Workflow Tested — ajrss2009 · 2026-08-08