Extend a photo on any side with FLUX.2 Klein 9B and Pixaroma's PixaOutpaint LoRA. Qwen3-VL describes the photo so the new area continues the same scene, and Stitch Inpaint restores every original pixel outside a 64 px seam.The first run extends the 3:2 golden-hour field photo from the sample inputs to a 4:5 portrait frame, top and bottom at once, at 1136 × 1424 on the 1.6 MP canvas.<h3 id="what's-inside">What's inside</h3><ul><li>Load Image + Pad: drag the borders outward on any side. Flat gray (#808080) fill, a 1.6 MP canvas and a 64 px seam feather.</li><li>Scene description: Qwen3-VL 8B describes the unpadded photo through the core Text Generate node. The prompt keeps the <code>pixaoutpaint</code> trigger and the "replace all solid gray areas" instruction first, then adds the description; Show Text displays what goes to Klein. The 8B model only captions; Klein keeps its own text encoder.</li><li>Reference and starting latent: the padded canvas is VAE-encoded once, attached as a reference latent to both conditioning branches, and sent to the sampler with the padding mask (Set Latent Noise Mask), so the known image anchors the result.</li><li>Sample: 4 steps, CFG 1, <code>euler</code> / <code>simple</code>, full denoise, fixed seed.</li><li>Stitch: Stitch Inpaint restores the source outside the 64 px feather, with tone match at 1. Compare slides the result against the padded input, and Save Image writes a PNG with the workflow embedded.</li></ul><h3 id="quick-start">Quick start</h3><ol><li>Update ComfyUI. The caption model needs a build with Qwen Image 2.1 support (mid-September 2026 or newer; I tested on 0.37.0).</li><li>Install AusBoss nodes from ComfyUI-Manager (search "AusBoss", registry id <code>ausboss-nodes</code>) and restart. You need 2.2.0 or newer, which is what Manager installs.</li><li>Unzip <code>sample_inputs.zip</code> into <code>ComfyUI/input/</code> (it holds <code>ausboss_golden_field_32.png</code>).</li><li>Drag the workflow in and download the five files from the card. The LoRA Loader holds PixaOutpaint at strength 1.</li><li>Queue. For your own photo, drag the pad edges on Load Image + Pad.</li></ol><h3 id="models">Models</h3><ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/Winnougan/Klein9b-Distilled-Base-INT8-Convrot/resolve/main/flux-2-klein-9b-int8-convrot.safetensors">flux-2-klein-9b-int8-convrot.safetensors</a> → <code>models/diffusion_models</code> · 9.43 GB</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/flux2-klein-9B/resolve/main/split_files/text_encoders/qwen_3_8b_fp8mixed.safetensors">qwen_3_8b_fp8mixed.safetensors</a> → <code>models/text_encoders</code> · 8.66 GB</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/flux2-klein-9B/resolve/main/split_files/vae/flux2-vae.safetensors">flux2-vae.safetensors</a> → <code>models/vae</code> · 336 MB</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Pixaroma/experimental_loras/resolve/main/pixaoutpaint.safetensors">pixaoutpaint.safetensors</a> → <code>models/loras</code> · 127 MB</li><li>caption only <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3vl_8b_int8_convrot.safetensors">qwen3vl_8b_int8_convrot.safetensors</a> → <code>models/text_encoders</code> · 9.35 GB</li></ul>The diffusion model is the INT8 build I tested, a community conversion by Winnougan. The Workflow Note asks for a distilled Klein 9B in that loader; Black Forest Labs' own files (<a target="_blank" rel="ugc" href="https://huggingface.co/black-forest-labs/FLUX.2-klein-9B">FLUX.2-klein-9B</a>, <a target="_blank" rel="ugc" href="https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-fp8">FLUX.2-klein-9b-fp8</a>) are gated on Hugging Face, so accept the license there first. The text encoder and VAE come from Comfy-Org's Klein 9B repo. The caption model uses CLIP type <code>qwen_image</code>; the renderer still uses Klein's own encoder.<h3 id="custom-nodes">Custom nodes</h3>Only <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss">ComfyUI-AusBoss</a>, 2.2.0 or newer (Manager installs it): free and open source. Everything else is core ComfyUI, including Text Generate for the caption. You don't need Pixaroma's nodes or an LLM pack.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. <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss#if-manager-cant-find-ausboss">See what to click</a>.<h3 id="settings-that-worked-and-what-didn't">Settings that worked, and what didn't</h3><ul><li>Keep the trigger and the instruction first: <code>pixaoutpaint, replace all solid gray areas with a seamless continuation of the surrounding scene.</code> In testing, other wording left the borders unfilled, even with the same seed. Scene details go after it; the caption stage adds them for you.</li><li>Flat gray fill (#808080) on Load Image + Pad.</li><li>Masked starting latent and a 64 px feather: at the same seed, this reduced the seam that an empty latent with a 24 px feather left, and the protected source interior stayed pixel-exact.</li><li>Tone match starts at 1; lower it if it shifts a new region too far. It blends a finished outpaint; it can't fill a region the model left gray.</li><li>If a border stays gray: check that the LoRA row is on and the trigger matches, restore the supplied prompt, try another seed, or extend fewer sides per pass.</li></ul><h3 id="speed">Speed</h3>Measured on an RTX 5090 (32 GB): 3 s for the sample, 3:2 to 4:5 on the 1.6 MP canvas, with the models loaded; 12 s on a first run right after another model.I haven't tested smaller cards yet. ComfyUI offloads what doesn't fit, so expect it to run slower rather than fail.<h3 id="about-the-gallery">About the gallery</h3>The PNG is a straight output of this workflow (seed 8311); drag it into ComfyUI to load its exact settings. The source photo was made with my Qwen Image 2.1 Studio workflow. The cover is cut from the real render.<h3 id="license-and-credits">License &amp; credits</h3>FLUX.2 [klein] 9B is by Black Forest Labs and released under the FLUX Non-Commercial License: non-commercial use of the model only (the 4B Klein models are Apache 2.0). The INT8 file is a community conversion by Winnougan; the text encoder and VAE are Comfy-Org's repackage. PixaOutpaint is Pixaroma's experimental Klein 9B outpaint LoRA (<a target="_blank" rel="ugc" href="https://huggingface.co/Pixaroma/experimental_loras">Pixaroma/experimental_loras</a>, MIT). The caption model is Qwen Image 2.1's Qwen3-VL 8B text encoder, repackaged by Comfy-Org, under the Qwen Research License (non-commercial use only). The workflow and the AusBoss nodes are free.<h3 id="changelog">Changelog</h3><ul><li>v1.0 (2026-09-25): first release.</li></ul>Questions or bugs: <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss/issues">GitHub issues</a> or the comments here. Post what you make with it; I read everything. I post new workflows on <a target="_blank" rel="ugc" href="https://x.com/Zanzibased">X @Zanzibased</a> and <a target="_blank" rel="ugc" href="https://github.com/ausboss">GitHub</a>.

Files with SHA256: f1528611b91b455ef840b9fa74ae5548d070b153c23bf3352031c48a20b23d6c