MultiDiffusion Panorama for FLUX β€” training-free ultra-wide generation (Modular Diffusers custom block)

πŸ€— Hub: remyxai/panorama-flux-modular Β· πŸ“„ Paper: arXiv:2302.08113 Β· πŸ’» Reference: omerbt/MultiDiffusion (unlicensed β€” see the clean-room note) Β· πŸ“¦ Monorepo: flux-recipes

Generate ultra-wide / panoramic images well beyond FLUX's native aspect β€” no training, no extra weights β€” as a Modular Diffusers custom block. Implements MultiDiffusion (arXiv:2302.08113): the wide latent is denoised through overlapping native-resolution windows, and each window's prediction is averaged back into the shared canvas every step, so the fused panorama is one globally coherent scene rather than stitched tiles.

a coherent 3:1 panorama

A single coherent wide scene β€” no repetition or seams at the window boundaries. Training-free.

Usage

import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("remyxai/panorama-flux-modular", trust_remote_code=True)
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

panorama = pipe(
    prompt="a sweeping mountain ridge at golden hour, alpine lake in the foreground",
    height=1024, width=3072,     # 3:1 β€” go to 4096 for a full wrap
    window=1024, stride=512,     # native tile + 50% overlap
    num_inference_steps=28,
).images[0]
panorama.save("panorama.png")

height/width are the full canvas; the window/stride knobs control the tiling. Everything must be a multiple of 16.

How it works

One wide latent is the canvas. At every denoise step the canvas is cut into overlapping windows and the FLUX transformer runs once per window; each window's velocity is scattered back into an accumulator and divided by how many windows covered that position (per-pixel mean β€” the MultiDiffusion fusion rule). The scheduler then advances the whole canvas once. Because windows overlap and are fused before the step, content crosses window boundaries freely: no visible seams, no repeated subject per tile.

Two details make it work on FLUX flow-matching:

  • Per-window img_ids. Each window gets native 2D position ids offset to its origin in the canvas, so local coordinates stay in-distribution while the offsets carry the global geometry.
  • The sigma schedule is computed at the window token count, not the canvas token count. Every window is native-resolution, so the dynamic-shift mu must be shifted by the window's sequence length.

This is exactly the failure the windowed design avoids: denoising a single 3072-wide latent stretches the position ids and the flow-match schedule far out of distribution, producing a smeared, textured "woven blob". MultiDiffusion keeps every window at native resolution β€” the schedule stays sane, and the image gets wide instead of mushy.

The transformer is never mutated (no attention-processor, hook, or pos_embed swap), so there is nothing to restore after a run. VRAM stays near a native 1024Γ—1024 generation per window; the wide latent is held once. The VAE decode uses tiling (a 4K-class decode otherwise dominates memory).

Key parameters

arg default meaning
width 3072 full canvas width (ultra-wide: 2048–4096; multiple of 16)
height 1024 full canvas height (multiple of 16)
window 1024 window side in px β€” keep at FLUX's native 1024
stride 512 window step in px; overlap = window - stride
window_height None window height in px; None = square windowΓ—window tiles
guidance_scale 3.5 FLUX guidance
num_inference_steps 28 denoise steps (applied to the whole canvas)

Tuning: the overlap is the quality lever. stride=512 (50% of a 1024 window) is a good, fast default; raise the overlap (lower the stride) if you see repeated content or a faint tonal band at a boundary β€” barely-overlapping windows repeat, and lightly-overlapping ones can leave a soft exposure step. Cost scales with the number of windows: a 3072Γ—1024 canvas at 1024/512 is 5 window passes per step, roughly 5Γ— a native 1024Γ—1024 step. For the smoothest result (the hero above): stride=256 (75% overlap) with guidance_scale=3.0 β€” the extra overlap plus slightly lower guidance removes the faint between-window banding, at ~2Γ— the passes.

Dependencies

diffusers (main / β‰₯ 0.40), transformers, accelerate, sentencepiece, protobuf. No extra weights, no fine-tuning, no preprocessing models.

Attribution & AI assistance

MultiDiffusion by Bar-Tal, Yariv, Lipman and Dekel. Clean-room implementation: the reference repo omerbt/MultiDiffusion carries no license, so none of its code was read or copied β€” this block was written from the paper's description alone (the method itself is not copyrightable). The same fusion rule is cited in our regional-prompting card.

Authored with AI assistance (Claude) and validated by the Remyx AI team; method credit to the MultiDiffusion authors. Uses FLUX.1-dev under its non-commercial license β€” this derivative inherits that restriction.

Citation

@misc{bartal2023multidiffusion,
  title={MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation},
  author={Bar-Tal, Dana and Yariv, Lior and Lipman, Yaron and Dekel, Tali},
  year={2023}, eprint={2302.08113}, archivePrefix={arXiv}, primaryClass={cs.CV},
  url={https://arxiv.org/abs/2302.08113}
}
Downloads last month
8
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including remyxai/panorama-flux-modular

Paper for remyxai/panorama-flux-modular