Nodes/ComfyUI-DiT360Plus/Kontext Panorama Editor
ComfyUI Node

Kontext Panorama Editor

FLUX Kontext, no inversion required

By thomashollier·Created 7 months ago·Updated 3 months ago· 3
Kontext Panorama Editor
  • pipeline
  • image
  • mask
  • IMAGE
prompt
steps28
guidance_scale3.5
strength1.00
seed0

If you try one editing workflow from this pack, make it this one. KontextPanoramaEditor repaints a masked region of a panorama using FLUX Kontext - Black Forest Labs' instruction-editing variant of FLUX.1 - and the whole pitch is that it skips the RF-inversion ceremony entirely. No inverter, no inversion steps, no tau/eta dance. Image in, mask in, prompt in, edited image out. That's it.

How it works

Under the hood it wraps the already-loaded pipeline into a FluxKontextInpaintPipeline, reusing the transformer, text encoders, VAE and scheduler already in memory - the source explicitly notes this costs no extra VRAM. One subtlety it handles for you: if the loader applied accelerate hooks for balanced or sequential offload, this node strips them first and reapplies diffusers' own offload, so you don't get double-hooking crashes. VAE tiling and slicing stay on. Then it just runs the Kontext inpaint pipeline with your mask, prompt, and settings.

The one thing you must get right upstream: the loader's base_pipeline has to be black-forest-labs/FLUX.1-Kontext-dev, and you want a Kontext-capable checkpoint in diffusion_models. The README also says to leave lora_id empty here - the DiT360 LoRA and the Kontext path are separate branches, don't mix them.

The inputs

  • pipeline - from Flux Panorama Loader (Kontext base).
  • image - the panorama to edit.
  • mask - white (1.0) = the area to repaint, black = preserve. No mode flip-flop to remember, which is the whole appeal.
  • prompt - what the masked region should become.
  • steps - 28 recommended. Kontext converges faster than the full DiT360 generation; don't feel the need to push it to 50.
  • guidance_scale - 3.5 recommended for Kontext.
  • strength - how much to modify the masked region. 1.0 (default) = full repaint; dial it toward 0.6–0.8 if you want the edit to lean on what's already there. This is the knob you'll tune most.
  • seed - reproducibility.

Output is an IMAGE at the same resolution as the input.

The workflow that makes it worth it

Direct panorama inpainting works, but Kontext editing a full 2:1 equirect image directly has a dirty secret: equirectangular projection distorts everything away from the horizontal centerline, and the model's edit quality suffers for it. The pack's flagship trick is to sidestep that:

Load Image (panorama) -> Equirect->Perspective -> Kontext Panorama Editor -> Perspective->Equirect

Extract a distortion-free perspective view with Equirect → Perspective, aim it at the region you care about, paint the mask on that clean view (the README recommends Impact Pack's PreviewBridge → MaskEditor for this), edit with Kontext, and stitch the patch back with Perspective → Equirect. You get edits that look like a normal photo instead of a funhouse mirror.

When you'd pick the other editor

The RF-inversion path (Image Inverter → Panorama Editor) gives you outpainting - extending the panorama past its edges - and finer source-fidelity control via tau/eta. Kontext can't outpaint, and its strength is a blunter instrument than the inversion knobs. But for 90% of "I want to change this thing in this panorama" jobs, Kontext is faster to set up, faster to run, and easier to debug. That's the one I'd reach for.

CategoryDiT360Plus/editing

Inputs (8)

NameTypeDefaultDescription
pipelineDIT360_PIPELINEFLUX pipeline from the Flux Panorama Loader (loaded with Kontext base_pipeline).
imageIMAGESource panoramic image to edit. (B,H,W,C) float32 [0,1].
maskMASKEditing mask. White (1.0) = area to repaint, Black (0.0) = preserve.
promptSTRINGDescribes the desired edit for the masked region.
stepsINT281–200Number of denoising steps. Recommended: 28.
guidance_scaleFLOAT3.50–20Classifier-free guidance scale. Recommended: 3.5 for Kontext.
strengthFLOAT1.000–1How much to modify the masked region. 1.0 = full repaint. Recommended: 1.0.
seedINT00–18446744073709550000Random seed for reproducibility.

Outputs (1)

NameTypeDescription
IMAGEIMAGE