Player not loading? Watch on YouTube
This Russian-language tutorial explains the speaker's generative upscaling workflows for ComfyUI, with separate SDXL and FLUX variants. The speaker prefers SDXL because it runs faster in their workflow and says FLUX uses more GPU memory, particularly during initial loading. These are observations about the demonstrated setup, rather than hardware benchmarks.
The workflow prepares an image, calculates tile dimensions for its aspect ratio, enlarges each tile and uses generation to add detail. Higher denoise can introduce more detail but also move the result further from the original. Hyper Caption supplies image information for tile generation; the speaker names Microsoft Florence as one possible recognition model. The SDXL walkthrough covers checkpoint selection, CFG and sampling steps, with 20 or 25 steps described as workable settings.
Overlap and padding help reduce visible joins between tiles. The example uses padding of 96, and the speaker recommends increasing it when seams appear, while noting that more padding can take longer. The workflow then assembles and saves the final image. The FLUX variant uses guidance and ControlNet to help preserve the image's structure.
The supplied setup notes point to ComfyUI-UltimateUpsacaler, with Windows batch commands and an optional FLUX asset installation flag. After installation, they instruct users to restart ComfyUI and load workflows\hyper_tile_sdxl.json. The speaker also says the custom nodes receive ongoing updates.
