Community Pipeline: Add HRDiT (training-free high-resolution FLUX.1-dev) - #26

Open

smellslikeml wants to merge 19 commits into

mainfrom

flux-hrdit-community-pipeline

Commits on Aug 14, 2026

Add training-free high-res Flux community pipeline (HRDiT)
github-actions[bot]
authored
5e013d1
fix(hrdit): add __call__ docstring so replace_example_docstring doesn't crash on import
smellslikeml
committed
3343901
fix(hrdit): initialize _joint_attention_kwargs before use (read-before-assign at inference)
smellslikeml
committed
075d542
fix(hrdit): use set_progress_bar_config for stage label (progress_bar takes no desc kwarg)
smellslikeml
committed
163b597
fix(hrdit): wrap __call__ in @torch.no_grad() and compile flex_attention (OOM: autograd graph + eager full-score materialization)
smellslikeml
committed
85435bb
fix(hrdit): faithful SPA re-port — monotonic bundle coarsening + in-attention variant averaging + proportional scale

smellslikeml
and
claude
committed
95e9d66
feat(hrdit): NTK-aware RoPE scaling + step-gated SPA (the actual high-res mechanism)

smellslikeml
and
claude
committed
cfd820e
feat(hrdit): high-fidelity port — frequency-domain structure guidance for the top stage

smellslikeml
and
claude
committed
5291b84
test(hrdit): update unit tests to current API (SPA variants, flux_rope NTK, structure-guidance helpers; drop HAP)
smellslikeml
committed
702d0c3
bench(hrdit): benchmark naive FluxPipeline vs HRDiT at target resolution (drop removed use_hap/group_num=1)
smellslikeml
committed
d738ed7
docs(hrdit): update community README to the NTK + SPA + structure-guidance method and simplified API
smellslikeml
committed
a413b38
docs(hrdit): add community pipeline overview-table row (with Colab)

committed
fe0a85f
style(hrdit): satisfy ruff — drop unused n_upscale, format exponent operators

committed
5f5fd01
style(hrdit): satisfy ruff — sort imports (I001), dict() -> literal (C408)

committed
1166542
style(hrdit): align copyright header with Flux community pipelines (Black Forest Labs + HuggingFace Team)

committed
9280328
style(hrdit): align test copyright header with diffusers convention

committed
6cdcf78
docs(hrdit): note MIT license of the reference implementation

committed
75cbf74

Navigation Menu

Community Pipeline: Add HRDiT (training-free high-resolution FLUX.1-dev) - #26

Community Pipeline: Add HRDiT (training-free high-resolution FLUX.1-dev)#26
smellslikeml wants to merge 19 commits into
mainfrom
flux-hrdit-community-pipeline

Commits on Aug 14, 2026

Add training-free high-res Flux community pipeline (HRDiT)

fix(hrdit): add call docstring so replace_example_docstring doesn't crash on import

fix(hrdit): initialize _joint_attention_kwargs before use (read-before-assign at inference)

fix(hrdit): use set_progress_bar_config for stage label (progress_bar takes no desc kwarg)

fix(hrdit): wrap call in @torch.no_grad() and compile flex_attention (OOM: autograd graph + eager full-score materialization)

fix(hrdit): faithful SPA re-port — monotonic bundle coarsening + in-attention variant averaging + proportional scale

feat(hrdit): NTK-aware RoPE scaling + step-gated SPA (the actual high-res mechanism)

feat(hrdit): high-fidelity port — frequency-domain structure guidance for the top stage

test(hrdit): update unit tests to current API (SPA variants, flux_rope NTK, structure-guidance helpers; drop HAP)

bench(hrdit): benchmark naive FluxPipeline vs HRDiT at target resolution (drop removed use_hap/group_num=1)

docs(hrdit): update community README to the NTK + SPA + structure-guidance method and simplified API

docs(hrdit): add community pipeline overview-table row (with Colab)

style(hrdit): satisfy ruff — drop unused n_upscale, format exponent operators

style(hrdit): satisfy ruff — sort imports (I001), dict() -> literal (C408)

style(hrdit): align copyright header with Flux community pipelines (Black Forest Labs + HuggingFace Team)

style(hrdit): align test copyright header with diffusers convention

docs(hrdit): note MIT license of the reference implementation

Commits on Aug 15, 2026

refactor(hrdit): delegate Euler update to scheduler.step; drop private scheduler API + dead code

docs(hrdit): update Colab validation notebook link