Give a clip a new shape: square to 9:16, vertical to 16:9, any side. An in-context LoRA for LTX-2.3 fills black bars with scene that moves with the footage, and Stitch Inpaint puts every original pixel back. Core LTX nodes plus my free AusBoss nodes; no LTXVideo pack.

The first run grows the square rooftop clip from the sample inputs into a 9:16 one, extending the sky above her and the rooftop below: 704 × 1280, 97 frames (about 4 seconds), in about 71 seconds on an RTX 5090, with the clip's audio kept.

What's inside

• Video Crop + Rotate + Pad → Clip: pick or upload a clip on the node. Tap the 9:16 chip under the preview: the whole clip stays, and centered black bands above and below bring the canvas to 9:16 (tap 16:9 for bands at the sides). Tap the lit chip again to lock the format, so the crop and padding handles keep it while you drag. Fill stays #000000 and feather 0.

• Trim: the rail under the preview is the timeline. Drag it to scrub, drag the IN/OUT handles to trim (the picture follows the handle), or type frame numbers. Snap 8n+1 keeps a frame count LTX accepts (49, 97, 121, 145), Limit 97 caps the sample at about four seconds, and Resize at 0.86 MP in steps of 32 lands the padded sample on 704 × 1280. Fixed frames gives you a window that slides without changing length.

• A prompt written for you: Select Frame takes the first frame from the clip node's original output, before any padding, and Generate Text runs on the same Gemma model the LTX text encoder loads. Its instruction asks for the whole finished shot, including what continues above and below. Show Text displays the result on its way into the positive encoder. Padding the sides instead? Change "a tall shot" to "a wide shot" and "continues above and below it" to "surrounds it to the sides".

• Sample: LTXV Add Guide and Get IC-LoRA Parameters carry the source frames in-context; 8 steps, CFG 1, euler_ancestral / linear_quadratic, with the distilled LoRA at 0.5 and the outpaint IC-LoRA at 1.0 on the dev transformer.

• Decode, stitch, save: a tiled decode, then Stitch Inpaint, fed by the clip node's stitcher, pastes the source frames back bit-for-bit and blends the new bands in over 32 px. Save Video muxes in the source audio.

Quick start

1. Update ComfyUI (I tested on 0.37.0).

2. Install AusBoss nodes from ComfyUI-Manager (search "AusBoss", registry id ausboss-nodes) and restart. You need 2.2.0 or newer, which is what Manager installs.

3. Unzip sample_inputs.zip into ComfyUI/input/ (it holds ausboss_day_sky_square.mp4), or pick your own clip on the Clip node.

4. Drag the workflow in and download the six files from the card. Close other GPU-heavy apps first.

5. Queue. For your own clip, tap a format chip (9:16, 16:9 and more), trim with the handles and check the prompt on Show Text.

Models

• ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors → models/diffusion_models · 23.5 GB

• gemma_3_12B_it_fp8_scaled.safetensors → models/text_encoders · 13.2 GB

• ltx-2.3_text_projection_bf16.safetensors → models/text_encoders · 2.31 GB

• LTX23_video_vae_bf16.safetensors → models/vae · 1.45 GB

• ltx-2.3-22b-distilled-1.1_lora-dynamic_fro09_avg_rank_111_bf16.safetensors → models/loras · 2.74 GB

• ltx-2.3-22b-ic-lora-outpaint.safetensors → models/loras · 1.31 GB

The projection file goes in the second slot of the dual text-encoder loader. The transformer, projection, VAE and distilled LoRA are Kijai's ComfyUI split of LTX-2.3 (Kijai/LTX2.3_comfy); an int8_convrot transformer in the same folder fits smaller cards.

Custom nodes

Only ComfyUI-AusBoss, 2.2.0 or newer (Manager installs it): free and open source. LTXV Add Guide, Get IC-LoRA Parameters, LTXV Crop Guides, Generate Text and the rest are core ComfyUI. You don't need the ComfyUI-LTXVideo pack or VideoHelperSuite.

Manager can't find AusBoss? In the classic Manager window, if the Channel box is empty, pick default, then click Install Missing Custom Nodes again. See what to click.

Settings that worked, and what didn't

I ran 53 renders over six clips: 640 × 360 to 1080p, 24 to 60 fps, 1.5 s to 108 s, phone footage, a TV sketch and model outputs. Two settings decided whether the bands got painted:

• Fill #000000. Pure black is the IC-LoRA's "paint here" signal. Gray, white or any color came back as a flat bar. #0a0a0a still painted; #202020 did not.

• Feather 0. A soft edge between footage and band hides the band from the model: feather 24 came back flat on every clip, feather 4 still painted.

Everything else painted: every format chip in both orientations, 2.5:1, one band only, all four sides, a 10° tilt, 0.5 to 1.3 MP, and 33 to 225 frames.

• Stitch color match at 0 for video. Its per-frame seam estimate can mistake moving objects for a tone shift and spread flickering bands across the new area.

• Dark backdrops get painted as more of the same, like a dark floor under a subject on a plain stage. That's the model reading the scene; a missed band is flat with a hard edge.

• Very dark footage (night, underwater) can read as a bar to the model. The LoRA author's fix is a gamma round trip: brighten (gamma 2.0) before the canvas, invert (gamma 0.5) after decoding.

• High frame rates: a 60 fps source with Limit 97 is only 1.6 s. Set Every nth to 2 or 3 to keep more seconds at 30 or 20 fps; both painted fine.

• Out of VRAM? Drop to 960 × 544, or swap the transformer for the int8_convrot build.

Speed

Measured on an RTX 5090 (32 GB) with the models loaded: 70.9 s for the sample, 97 frames at 704 × 1280. In an earlier test at the same size, VRAM peaked at 31.9 GB, and a 225-frame clip at 1280 × 704 took about three minutes.

I haven't tested smaller cards yet.

About the gallery

Both clips are straight outputs of this workflow from the same square source, with its audio: the 9:16 default (704 × 1280, 97 frames) and the same clip with the 16:9 chip (1280 × 704). The source is a still from my Qwen Image 2.1 Text to Image workflow, animated with my MiniMax H3 Image to Video workflow. The cover is cut from the real clips: the square source, the black padding growing above and below it, then a wipe to the result. There's a still of it too.

License & credits

LTX-2.3 is by Lightricks, released under the LTX-2 Community License: entities with US$10M or more in annual revenue need a paid commercial license from Lightricks. The ComfyUI split files are by Kijai. The Gemma 3 12B text encoder is Google's model, repackaged by Comfy-Org in its LTX-2 repo. The outpaint IC-LoRA is by oumoumad, under the LTX-2 Community License. The sample clip was animated with MiniMax H3 from a Qwen Image 2.1 still, so the MiniMax H3 Community License covers that clip (read it on the MiniMax-H3 model page), and Qwen Image 2.1 is non-commercial under the Qwen Research License. The workflow and the AusBoss nodes are free.

Changelog

• v1.0 (2026-09-25): first release.

Questions or bugs: GitHub issues or the comments here. Post what you make with it; I read everything. I post new workflows on X @Zanzibased and GitHub.