360 Panorama: Equirectangular Projection for Krea 2

作者：Azimondious

I asked plain Krea 2 for 360 panoramas and got wide-angle shots with mirrored repeats, stitched end to end. My test kitchen came out with its window wall on both sides of the picture.

This LoRA draws the projection: one continuous room or street around the viewer, floorboards curving, the ceiling or the sky stretching toward the top edge, the camera at eye height and the horizon near the middle. I made it for skyboxes, VR backdrops and 360 viewers. There is no trigger word. The phrase a 360 degree equirectangular panorama of opens the prompt.

Open the finished panorama in a 360 viewer and you can turn round inside one kitchen.

• The projection is what it adds. With the LoRA that kitchen came out as one room around the camera, and a pine forest wrapped with no break I could see.

• The seam needs the workflow. On its own the LoRA leaves a line where the left edge meets the right, about as often as plain Krea 2 does. My free Krea 2 Seamless Tile workflow closes it: after its panorama graph, 12 of 12 test renders passed my seam check.

HOW TO USE

• Load it in ComfyUI with LoraLoaderModelOnly at strength 1.0.

• There is no trigger word. Start the prompt with a 360 degree equirectangular panorama of, then the place in plain words. Every training caption opens this way.

• Render at 1408x704, exact 2:1. My training images and test renders are all that size.

• Then run the Seamless Tile panorama graph on the result. Load the render in its blue node, set the red number to 704 (half of 1408), paste the same prompt into the green node and run. I tested its defaults, denoise 0.85 and band 256.

• The graph's preview puts the panorama next to itself. Look at the middle of it. If a line shows, change the seed and run again.

Example prompt at 1.0 (1408x704, seed 9151). It is the first showcase image, after the graph:

a 360 degree equirectangular panorama of a busy old town street on a rainy evening Example long prompt at 1.0 (1408x704, seeds 9151 and 9152). Half the training captions are a paragraph like this one and close on the same sentence about the sphere. Indoor captions say ceiling and floor where outdoor ones say sky and ground:

a 360 degree equirectangular panorama of a big old railway station hall. Iron arches hold a glass roof, long wooden benches stand on a stone floor, and cold daylight falls from above. The view is a full sphere unwrapped into one 2:1 image, the horizon level across the middle of the frame, the ceiling filling the top and the floor the bottom. I scored the two long test paragraphs by their seam number and have not judged their projection by eye. The short form is the one I have looked at.

IF IT GOES WRONG

A line where the edges meet. Expect it. The LoRA does not close the wrap. Run the panorama graph, and if a line survives the graph, give it another seed.

Something at the edge changed after the graph. The graph repaints a band over the join, so whatever stood there can turn into something else. In the workflow's own test a tree cluster became a sunlit ridge. Another seed gives you another band.

The top and bottom look smeared. Those are the poles. A real equirectangular image stretches its top and bottom rows, and the 60 HDRIs I trained on have the same stretch. The graph leaves the poles alone.

You asked for an anime panorama and got a photograph. It can happen. I trained on photographs. My one drawn test prompt kept its anime look, and one image is too few to promise that yours will.

RECOMMENDED SETTINGS

• Checkpoint: RedCraft 3.0 (Krea 2). The one I tested.

• 8 steps, CFG 1, sampler er_sde, scheduler simple.

• 1408x704, exact 2:1.

• Strength 1.0, the one strength I rendered. If you try 0.8 or 1.2, tell me what it did.

• Then the Seamless Tile panorama graph: denoise 0.85, band 256, red number 704.

• Other checkpoints and other UIs are untested.

WHAT I TESTED

Twelve test prompts, 2 seeds each, on steps 1500, 1750 and 2000, next to the same prompts and seeds with no LoRA. Six are short scenes: a pine forest, a tropical beach, an old town street in the rain, a desert canyon, a kitchen and a snowy village at night. Two are long paragraphs, a riverbank and the station hall above. Three use the wording I tried on plain Krea 2 before training (an alpine lake, a library, a rooftop garden). The last one has no phrase at all, an ordinary 1152x896 photograph of a lake.

My seam number is the difference between the first and last column over the median difference between neighbouring columns. Under 1.5 passes. All 60 training HDRIs pass it.

• The LoRA alone: at most 2 of the 18 short and reworded renders passed, on any step. Plain Krea 2 passed 2.

• Through the panorama graph: I ran the six short scenes, both seeds, at the graph's defaults. Steps 1750 and 2000 each passed 12 of 12. Plain Krea 2 passed 11 of 12 through the same graph, so the graph closes the seam with or without the LoRA. It repaints a band over the join and leaves the rest of the picture untouched, so a stitched layout stays stitched.

• By eye, the seam: I turned four finished panoramas half-way round so the join sat in the middle (two kitchens, the street, the beach). I found no seam and no band on any of them.

• By eye, the projection: the forest, the street and the kitchen side by side with plain Krea 2. The kitchen is the clearest: plain Krea 2 repeats its window wall, and the LoRA draws one room with curved floorboards and a ceiling that stretches toward the top. I cut ordinary 90-degree camera views out of the finished kitchens, as a 360 viewer does: plain Krea 2 showed me the same stove in front and behind, and the LoRA showed four different walls, a floor where the cabinets meet at a corner and a ceiling with lights.

• No phrase: loaded with the photograph prompt, it drew an ordinary photograph on both seeds.

Two seeds per prompt is a small sample, so read the counts as a ranking.

LIMITS

• On its own it leaves a seam at the wrap, about as often as plain Krea 2. Krea 2 has no wrap-around padding, so 60 photographs taught the LoRA the look, and the graph handles the edge rule.

• The graph repaints a band over the join, so what stood at the edges can change.

• The poles stretch, as in any equirectangular image, and the graph does not touch them.

• All 60 training images are real photographs. My one drawn test prompt, an anime rooftop garden at night on one seed, kept its drawn look on step 2000, the published file. That is a single image, so drawn panoramas are still barely tested.

• Tested at 1408x704 and strength 1.0. Other sizes and strengths are untested.

• I scored the long paragraphs by seam number alone.

• I checked the horizon by eye on a handful of renders. It sat near the middle on those.

• I trained one wording of the phrase. I have not judged other wordings by eye.

• Stacking it with other LoRAs is untested.

• Tested on RedCraft 3.0 alone.

MAKER'S NOTE

Trained on Krea 2 Raw with ai-toolkit (arch krea2), rank 32, batch 1, lr 2e-4, about 1 MP buckets, 2000 steps. It runs on RedCraft. The file is step 2000, the final step.

The dataset is 60 HDRIs from Poly Haven, all CC0. 40 are outdoor: streets and squares, rooftops, beaches, harbours, forests, desert, six at night and three in snow. 20 are indoor: halls, museums, a theatre, a church, a kitchen, workshops, an indoor pool. I resized each to 1408x704 and kept the ones whose own left and right edges pass the seam number. Seven failed and went out. Indoor HDRIs failed more often, on hard vertical edges where the stitch closes.

Each image is in the dataset twice, once with a short caption and once with a paragraph that ends on the sphere sentence, 120 pairs. The captions carry no place names and no brand names.

Thanks to Poly Haven and its photographers for putting this library out under CC0.

The companion is Krea 2 Seamless Tile, free, core nodes only. Its sibling is Camera Angles, which moves the camera up and down. I have not tested the two together.

If this worked for you, a thumbs-up here helps the next person find it.

Free to use. Credit is appreciated. The licence is on this page.