DyPE for FLUX — training-free ultra-high-resolution (Modular Diffusers custom block)

Training-free text-to-image up to 4096² from off-the-shelf FLUX.1-Krea-dev — no fine-tuning, no extra weights, single pass — packaged as a Modular Diffusers custom block. Load it and generate in three lines.

_{4096² from stock FLUX.1-Krea-dev, training-free, method="spectral".}

Why training-free high-res?

Ask stock FLUX for a 4096² image directly and it collapses — at 65k tokens the flow-match sigma schedule blows up and the denoiser never denoises (a "woven blob"). DyPE (arXiv:2510.20766) fixes this at inference, in a single pass: a timestep-dynamic YaRN/NTK-by-parts RoPE schedule (strength κ=t², strongest early when global structure is set, fading as detail resolves) plus the flow-match shift cap that keeps the schedule sane at high token counts. No ladder, no fine-tuning.

Usage

import torch
from diffusers import ModularPipeline

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

image = pipe(
    prompt="a clear blue sky over a calm turquoise sea, a lone red sailboat",
    height=4096, width=4096,
    guidance_scale=4.5,
    method="spectral",          # "yarn" = DyPE · "spectral" = DyPE + SEGA (less speckle)
).images[0]
image.save("dype_4k.png")

The FLUX.1-Krea-dev transformer / VAE / text-encoders stream from the base repo — nothing is duplicated here. VAE tiling is enabled automatically (at 4K the VAE encode/decode dominates memory).

The speckle fix: `yarn` vs `spectral` (SEGA)

DyPE alone leaves a faint high-frequency speckle in flat regions — a genuine artifact of its scalar attention-temperature at high resolution. method="spectral" swaps that scalar for a per-frequency, content-aware attention temperature derived from the latent's Fourier spectrum each step (SEGA, arXiv:2605.22668) — it suppresses the high-frequency band selectively, so the speckle goes without softening real detail.

Same flat water region, native resolution — the woven speckle (left) is gone under spectral (right):

Because SEGA modulates attention every step, it also shifts composition at a fixed seed — it is an alternate mode, not a pixel-match of yarn. Use spectral for clean flat regions (skies, water, walls); yarn reproduces plain DyPE.

How it works

DyPE is applied by swapping the transformer's positional-embedding module for a dynamic one and feeding it the current timestep (and, in spectral mode, the running latent's spectral profile) via a native forward pre-hook — no changes to the base weights, restored on exit.

Dynamic YaRN / NTK-by-parts RoPE — extends the trained position band to the high-res grid, scaled by κ=t² so extrapolation is strongest early and fades as the image resolves (the core DyPE mechanism).
Flow-match shift cap — caps the schedule shift (mu = max_shift) so the sigma schedule stays sane at 65k tokens; required for any single-pass >2K FLUX generation.
SEGA spectral mode (optional) — per-RoPE-dimension attention temperature from the latent's FFT energy profile; the selective high-frequency fix above.

Unlike a resolution-ladder approach, DyPE is single-pass — one denoise at the target resolution. DyPE's RoPE schedule in this block is verified bit-exact against the reference implementation; the pipeline produces coherent 4K where stock FLUX collapses, and spectral suppresses the speckle (shown above).

Key parameters

arg	default	meaning
`height`, `width`	1024	target resolution (multiples of 16); DyPE engages above 1024²
`method`	`"yarn"`	`"yarn"` = DyPE · `"spectral"` = DyPE + SEGA
`guidance_scale`	4.5	FLUX.1-Krea-dev's realism range
`num_inference_steps`	28	denoise steps
`dype`	`True`	set `False` to A/B against the (blob-prone) stock path

Running on less VRAM

A 4096² generation peaks ~49.8 GB (bf16) with VAE tiling on an A100. The same 4-bit recipe proven for the sibling HRDiT block applies here — quantize the transformer + T5 to NF4 and swap them in with update_components — which brought that pipeline to ~29 GB @ 4096² / ~16 GB @ 2048²; the DyPE-specific measurement is a pending follow-up. At any resolution, keep pipe.vae.enable_tiling() (on by default here).

Attribution & AI assistance

Training-free reimplementation for Modular Diffusers of DyPE (MIT, guyyariv/DyPE); the spectral mode adapts SEGA from wildminder/ComfyUI-DyPE (Apache-2.0). The Modular-Diffusers adaptation was authored with AI assistance (Claude) and reviewed + validated by the Remyx AI team; all credit for the methods goes to the original authors.

Citation

@misc{issachar2026dype,
  title={DyPE: Dynamic Position Extrapolation for Ultra High Resolution Diffusion},
  author={Issachar, Noam and Yariv, Guy and Benaim, Sagie and Adi, Yossi and Lischinski, Dani and Fattal, Raanan},
  year={2026},
  eprint={2510.20766},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2510.20766}
}