Skip to main content
Two complementary routes to 4K generation from off-the-shelf FLUX checkpoints, with no fine-tuning and no new weights, surfaced from arXiv, prototyped by Outrider, and taken toward huggingface/diffusers. The pair puts the same problem, training-free high resolution on a RoPE DiT, into two different library shapes. The choice of shape is itself the contribution.
Status as of November 2026. The linked PRs and Issues are the live record.

​
1. HRDiT: community pipeline, closed unmerged

4K (up to 4096²) from off-the-shelf FLUX.1-dev, with no fine-tuning, via NTK-aware RoPE scaling, monotonic bundle coarsening (SPA), and a structure-guided progressive ladder. Users get 4K straight from the base checkpoint instead of the generate-then-upscale or train-a-high-res-model workarounds. Submitted as a self-contained community pipeline (examples/community/pipeline_flux_hrdit.py) with no core-library changes. Built minimal-first: a position-alignment port established parity, then NTK RoPE scaling and structure guidance were layered onto only the stages that needed them. The speed-oriented HAP and DWT stages were documented and deferred, not shipped.

​
2. DyPE (plus optional spectral attention): hook, proposed

apply_dype(pipe.transformer) swaps the transformer’s rotary embedding for a timestep-aware YaRN / NTK-by-parts schedule (method="yarn"), a no-op at or below the trained resolution. method="spectral" adds a per-frequency, content-aware attention temperature computed from the latent’s spectral energy at each denoising step. Validated on FLUX.1-Krea-dev at 4096²:
  • method="yarn" reproduces the reference implementation (guyyariv/DyPE, MIT) bit for bit (Δ=0 on the positional-embedding output).
  • method="spectral" cuts flat-sky high-frequency speckle about 6× (Laplacian variance 76.5 → 12.7) while keeping fine detail, a separation a single scalar attention temperature can’t achieve.
Sky speckle comparison between the yarn and spectral methods

DyPE spectral attention: sky speckle, yarn vs spectral

State (the timestep, and for spectral the latent’s spectral profiles) is fed through a native register_forward_pre_hook, so it survives enable_model_cpu_offload. The spectral mode implements SEGA but is exposed as method="spectral" to avoid colliding with the existing Semantic-Guidance “SEGA” in diffusers.

​
Calibrated shape: the real contribution

HRDiT self-contains cleanly as a community pipeline. DyPE, by contrast, modifies a model internal (the positional embedding) at inference, the same surface as diffusers’ core hooks (apply_faster_cache, apply_pyramid_attention_broadcast). But training-free high-resolution methods have also historically landed as community pipelines (DemoFusion). Precedent points both ways, so the placement is raised with the maintainers in #14520 before opening the PR rather than assumed. A pipeline-level caveat surfaced during validation: above about 2K, FLUX’s default flow-matching shift (mu) grows with sequence length and collapses the sigma schedule, so both methods pin base_shift == max_shift. It’s documented as a usage step, since a positional-embedding hook can’t reach the pipeline’s scheduler.

​
Outrider’s role

Both methods were surfaced from arXiv and drafted by Outrider; DyPE through a brief-anchored run. The draft is the start. Parity checking, the spectral port, the shift and offload fixes, the rebase onto current main, and the maintainer-facing framing were the substantial follow-on. What Outrider removes is the cold start: turning “this paper looks relevant” into a wired, testable branch to harden.

​
References