pano_agent

Tools for generating 360° equirectangular panoramas from a single perspective image, plus a ComfyUI workflow builder for multi-view interior rendering.

---

Modal panorama scripts

Three GPU scripts that run on Modal, each implementing a different outpainting approach. All three share the same first step: DepthPro estimates the camera's focal length so the perspective image is projected onto the equirectangular canvas with geometrically correct FOV before any model fills in the unseen regions.

Source image

Single perspective image of a vintage trailer interior — the input to all three pipelines.

---

WorldGen (worldgen_pano.py)

Uses FLUX.1-dev Fill + WorldGen's panoramic LoRA. Inpaints the unseen ~75% of the sphere after projecting the reference image onto a 1024×512 canvas.

modal run worldgen_pano.py --image input.jpg --prompt "vintage trailer interior"

Base model	FLUX.1-dev Fill
LoRA	WorldGen panoramic (worldgen_img2scene.safetensors)
Output	1024×512
GPU	A100 40GB

---

DiT360 (dit360_pano.py)

Uses FLUX.1-dev + DiT360's panoramic LoRA with circular-padded attention and RFID inversion (Personalize Anything) to preserve projected pixels while outpainting. Higher fidelity at full panorama resolution.

modal run dit360_pano.py --image input.jpg --prompt "vintage trailer interior"
# Optional: --tau 30  (lower = tighter lock to source image, default 50)

Base model	FLUX.1-dev
LoRA	DiT360 (Insta360-Research/DiT360-Panorama-Image-Generation)
Output	2048×1024
GPU	A100-80GB (~37 GB VRAM)
Paper	arxiv 2510.11712

---

HunyuanWorld (hunyuan_pano.py)

Uses FLUX.1-Fill-dev + HunyuanWorld's panoramic LoRA. Best results for preserving interior geometry — narrow spaces stay narrow. Supports higher resolutions and more inference steps for finer detail.

modal run hunyuan_pano.py --image input.jpg --prompt "vintage trailer interior" \
  --pano-h 1024 --pano-w 2048 --steps 75

Base model	FLUX.1-Fill-dev
LoRA	HunyuanWorld-PanoDiT-Image (tencent/HunyuanWorld-1)
Output	1024×2048 (configurable)
GPU	A100 40GB
Repo	HunyuanWorld-1.0

---

Setup (one-time)

pip install modal
modal setup
modal secret create huggingface HF_TOKEN=hf_yourtoken

---

ComfyUI workflow builder (pano_agent)

Generates multi-view flat-on cardinal renders of an interior space from a single reference image, emitting a ComfyUI workflow JSON.

pano_agent/
├── pano_agent/        ← shared library (brief, build, prompts)
├── cli/               ← Python CLI pipeline (one-shot, scriptable)
├── claude_code/       ← Claude Code pipeline (interactive, iterative)
└── tests/             ← test suite for the shared library

Quick start — CLI

cd cli/
pip install anthropic
export ANTHROPIC_API_KEY=sk-...
python pano_agent_cli.py run ../reference.png -o ../workflow.json

Quick start — Claude Code

cd claude_code/
claude --append-system-prompt "$(cat SESSION.md)" \
       "Analyze this reference image and write scene_brief.json. <attach reference.png>"
python build.py scene_brief.json -o workflow.json

The build pass generates 8 views (4 cardinals + 4 interstitial corners) with prompts that enforce flat-on camera, world-space lighting, canonical window specs, and no corner recesses. See pano_agent/prompts.py.

Tests

python -m pytest tests/ -v