comfyUI-JigSaw-Hi-Res

🧩 Jigsaw Hi-Res

Generate beyond your model's native resolution — without the repeating-pattern artifacts.

A ComfyUI node that patches a Flux/Krea2-family model's rotary positional embedder (RoPE) with YaRN and NTK-aware frequency scaling, the same technique used to extend context length in LLMs — applied here to 2D image positions instead of 1D token sequences.

The problem

Flux/Krea2-style diffusion transformers encode spatial position with RoPE, tuned for a specific training resolution (commonly ~1024×1024). Ask for a significantly larger image and the position embeddings extrapolate past what the model has ever seen — the usual symptom is duplicated limbs/faces, tiled repetition, or a hazy "gray veil" wash over the image.

What Jigsaw Hi-Res does

At apply time, it:

1. Reads your target resolution (explicit width/height, or auto-detected from a connected latent).
2. Computes an adaptive YaRN ratio from how far the target resolution sits above the ~1024px reference (blending a linear and an area-based ratio, softly capped at 1.25× to keep contrast stable at extreme resolutions/aspect ratios — this cap is what keeps the classic "gray veil" oversaturation from creeping in on the more aggressive high-res attempts).
3. Swaps the model's pe_embedder for a custom RoPE implementation that reproduces comfy.ldm.flux.math.rope()'s exact output format ([..., N, dim/2, 2, 2] rotation matrices), but computes each frequency band's scaling individually — near-DC frequencies stay untouched (extrapolated), high frequencies get interpolated/compressed, following the standard YaRN "NTK-by-parts" ramp (beta_fast/beta_slow cutoffs).
4. Does all of this without retraining or fine-tuning — it's a pure inference-time patch on the position embedder.

Installation

Drop the folder into ComfyUI/custom_nodes/, restart ComfyUI. The node appears under Jigsaw/Patches → 🧩 Jigsaw Hi-Res.

Usage

Connect your MODEL in, plug a LATENT in (optional — auto-detects the real target resolution from it) or set width/height explicitly, and sample as usual at your target resolution.

Parameters

| Parameter | Default | Description | |---|---|---| | width / height | 1024 / 1024 | Target resolution. Overridden automatically if latent is connected. | | method | yarn | yarn — per-frequency-band interpolation (recommended, best quality at large upscales). ntk — single global frequency-base rescale (cheaper, less precise). base — disables scaling entirely (pass-through RoPE). | | rope_percentage | 1.0 | Blend factor for how much of the computed scaling to actually apply (1.0 = full effect). | | dype_enabled | True | Master switch for the whole dynamic scaling path. Off = falls back to the model's original, untouched RoPE. | | max_pe_length | 64 | (currently reserved — not yet wired into the scaling math) | | latent (optional) | — | If connected, its shape is used to auto-derive the true target width/height (×8 VAE factor) instead of the manual width/height inputs. |

Compatibility

Requires a model whose diffusion backbone exposes a pe_embedder module with .theta and .axes_dim attributes, and whose native RoPE kernel consumes the same [..., dim/2, 2, 2] rotation-matrix format as comfy.ldm.flux.math.rope() (Flux.1 and Krea2/Flux.2-Klein-family models). Models using a different RoPE convention are not supported and will raise a clear error on load rather than silently producing garbage.

Credits

By Jigsaw & Zeus — zeus.onl