ComfyUI
Stefan argues that you can skip slow ray-traced rendering. Export a deliberately flat, shadowless, high-contrast image from Blender and run it through a local, free ComfyUI SDXL image-to-image pipeline (caption, LLM-rewritten prompt, depth/canny ControlNets, upscale). He says this gives a photorealistic result in about 10 seconds, or 3-4 seconds on a very good GPU, which he calls roughly 100 times faster.
The video's core move is to split the labour. Blender only supplies the structure and colours of the scene, with shadows switched off on purpose and contrast raised so separate objects can be told apart. The AI pass supplies lighting, materials and realism.
Colour stays under the artist's control in Blender because the pipeline carries colour information through several channels at once: the caption, the rewrite instruction to 'keep the important information about the colors', and the image-to-image latent. Structure comes from the ControlNet maps. Stefan says the redundancy is deliberate: 'Never Enough we're working with AI'.
A raw caption of a flat, shadowless scene would only describe a flat scene. So an LLM rewrites the caption, in the way he compares to ChatGPT or Claude, into an attractive description with nice lighting and a landscape behind the windows. Colours are preserved, and style presets are added on top, so the diffusion model is steered towards a polished render.
The approach is presented as general, not limited to interiors. A screenshot of a flower vase works too, provided the prompt-editing step is adapted to the new subject. Results are impressive, with the cotton 'looks so real', but imperfect: the floor information was missed and the model took too much freedom with colours, which he says denoise tuning might fix.
Stefan picked SDXL for its ecosystem rather than for being the newest model. Many LoRAs and ControlNets already exist for it, and add-ons must be built for the specific base model. He calls Flux and SD 3.5 promising, but says Flux has far fewer add-ons for now.
The workflow is meant to be practical to adopt. He shares it in Notion with links to custom nodes and models, installs most nodes through the ComfyUI Manager, and says a few nodes (Florence and the LLM node) may need manual installation from GitHub. Speed depends on the user's GPU, and he frames the whole thing as a worthwhile, free, local investment of time.
Flat, shadowless, high-contrast Blender base render — A deliberately minimal export from Blender with all shadows disabled and very high contrast, so separate objects can be distinguished and used as the structural and colour input for the AI. Apply: Turn off shadows in the Blender scene, boost contrast, set the object colours you want, and export the flat image as the workflow's image input.
Florence-2 detailed captioning — A Florence model loaded through a custom-node download-and-load node that is tasked with producing a more detailed caption of the input image. Apply: Connect the image to the Florence caption node, set the task to detailed caption, and use the output description as the raw text for the next step.
LLM prompt editing with Searge LLM — A step that sends the caption to a local language model (transcribed as 'search LM' with a '7B' model) with an instruction to rewrite it, working like ChatGPT or Claude. Apply: Instruct it to treat the input as an image description and make it sound like a really nice render with good lighting and a landscape behind the windows, while keeping the colour information; edit the instruction for non-interior subjects.
SDXL Prompt Styler — A custom node that edits the prompt with preset styles, replacing hand-typed phrases such as 'Photo realistic'. Apply: Pick a preset from the list, in his case a realistic one (transcribed 'R style HP realis'), to format the prompt for the SDXL diffusion model.
Input image preparation (resize plus brightness/contrast) — A small image-processing block that sets the resolution SDXL expects and slightly adjusts brightness and contrast, replacing Photoshop touch-ups. Apply: Set the resolution in the resize node and tweak brightness and contrast slightly to suit your input, for example a SketchUp screenshot.
Upscale-then-resize-back detail pass — An optional step that upscales the input image and then resizes it back, which Stefan found keeps as much detail as possible. Apply: Run the prepared input through an upscale and then a resize back to the working resolution before the sampling stage.
Juggernaut XL checkpoint with image-to-image generation — The main model block: Juggernaut XL (a Stable Diffusion XL 1.0 model from CivitAI) outputs a model, a CLIP and a VAE, with the CLIP split into positive and negative prompts and the K sampler producing the image. Apply: Load the checkpoint, feed the positive prompt from the prompt chain, pass the model through the LoRA and ControlNet blocks to the K sampler, then decode the result.
ControlNet depth and canny — ControlNets that give control over what is generated, using a depth map and a canny outline map created from the input image with a ControlNet preprocessor node. Apply: Create depth and canny images from the input, check that the ControlNet type is supported by SDXL, and raise the strength if the output does not follow the lines accurately enough.
CR Apply Multi-ControlNet stack — A node that combines several ControlNets, with a switch and a strength setting for each, and a third slot that can be left off. Apply: Switch on the depth and canny entries, set each strength to a moderate value, and adjust upward for more accuracy, noting that 100% does not give very good results.
K sampler settings: CFG, steps, seed — The sampling controls: CFG governs how closely the prompt is followed and is taken from the model's recommended settings, steps is the number of iterations, and seed fixes the randomness. Apply: Leave CFG at the model's recommended value, experiment with steps, and keep the seed fixed when you want identical results from an unchanged input.
Denoise tuning for image-to-image — Image-to-image works by covering the input latent with noise, and the denoise level sets how much noise is added: more noise recognises less of the original, and less noise stays closer to it. Apply: Experiment within the range he gives (transcribed '1.8 and 1.95'), settling on about the value he transcribes as '92' to carry colour information on top of the ControlNets, and lower it if the model takes too much freedom with colours.
LoRA (low rank adaptation) — An add-on trained on top of the main model to improve a specific task, such as nice interiors or clouds, which also adds extra prompting. Apply: Choose a LoRA made for the same base model (SDXL 1.0), read its description to see what it improves, and load it between the checkpoint and the sampler.
Final upscale with an upscale model — A last step that loads an upscale model from the ComfyUI Manager and applies it to the generated image. Apply: Load the upscale model, run Upscale Image with it, and use the 0.5 setting that the video says roughly doubles the resolution.
Base-model compatibility check via CivitAI — The rule that LoRAs and ControlNets must be developed for the same main model as the checkpoint, with CivitAI named as the source for finding them. Apply: Before downloading a LoRA or ControlNet, confirm it matches the base model, here SDXL, and take any extra models from CivitAI.
Preview nodes for intermediate results — Black preview nodes in the graph that show intermediate results such as the ControlNet maps and the base generation. Apply: Watch them during a run to see where in the pipeline something went wrong, for example a missed floor.
Installing the workflow via ComfyUI Manager — The setup route: install ComfyUI Manager, open the provided workflow, use 'install missing custom nodes' (about 80% of nodes), then restart ComfyUI. Apply: After the automatic install, handle Florence and the Searge LLM node manually from their GitHub instructions if needed, and download the 7B model from the provided link or GitHub.
The shortcut works because the Blender export is only a blueprint. Shadows are disabled and contrast is high so objects separate cleanly. The AI is then responsible for all the lighting and surface detail that would normally cost render time.
The captioning model's direct description is 'not really what I need'. The extra rewrite step exists to add 'real nice lightning' and a nice landscape while still keeping the colour information, so colours can be set in Blender and survive into the render.
Stefan found by experimenting that upscaling and then resizing back to the working size helps keep as much detail as possible. He marks this step as optional.
Because SDXL is strict about input resolution, the input image is resized first, with a slight brightness and contrast adjustment. He says this removes the need for Photoshop and that a screenshot from SketchUp could be used as input.
He says not every ControlNet type is supported by SDXL, so you have to check which ones are. He uses depth and canny, the latter giving an outlined version of the image. If the model does not follow the lines accurately enough, you can raise the strength, but he says setting it to 100% 'not really gives us very good result'.
CFG is taken from the base model's recommended settings, and he says changing it makes the result worse. The knobs worth playing with are the number of steps and the denoise level.
Denoise is the trade-off between faithfulness and freedom. More noise means the model recognises less of the original image, and less noise keeps the output close to it or even unchanged. Because depth and lines already come from the ControlNets, he still wants image-to-image to carry colour. The transcript gives his explored range as '1.8 and 1.95' and his preferred value as 'something like 92'.
A same seed with an unchanged input gives exactly the same result, which makes tuning other settings repeatable.
A LoRA ('low rank adaptation') adds specialised training on top of the main model. He uses an interior LoRA made for SDXL 1.0 and gives clouds as an example of a weakness a LoRA could fix. LoRAs and ControlNets must be built for the same base model as the checkpoint.
For the flower test the rewrite prompt, which was tuned for real-estate interiors, had to be edited. He removed the furniture mention and asked for a photorealistic, awesome render, because the output otherwise described a living room.
The black nodes in the graph are preview nodes that show intermediate results such as the ControlNet maps and the base generation.
The final upscale model is loaded from the ComfyUI Manager. A setting of 0.5 is described as increasing the resolution about two times.
«how using AI render anything 100 times faster and absolutely for free using just your computer»
— 00:00
«disabled all the Shadows the only thing here and you will later know why is a contrast I just added a very high contrast so we can actually distinguish the separate objects»
— 01:27
«think of it as CH GPT or like CLA basically the same mechanism right»
— 03:47
«sdxl model we're using is quite strict about the resolution you put inside right»
— 05:30
«there is some compatibility uh issues you need to take into account like for example if you use sdxl there are some additional things like low and control net which we will cover just in a moment they need to be actually developed for that main model»
— 07:25
«whenever you have the same value the result will be exactly the same if the input is not changed right»
— 11:09
«the more noise the less it will recognize the original image and take some details from there»
— 11:34
«Laura which stands for low rank adaptation so Laura is something you can use along with main model that will improve your results by teaching it on top right to do some specific task»
— 12:45
«oh my God I mean this is actually more beautiful than I actually imagined»
— 16:53
«I use sdxl here because this it has a lot of control net stuff relatively old so a lot of lures and control net exist for it»
— 18:02
Reception
Viewers are broadly appreciative and engaged with the workflow, praising it and asking about extensions like video, animation and Flux, though a recurring thread of installation and troubleshooting complaints, mainly the Searge LLM node, tempers the positivity.
This is a concrete, reproducible walkthrough of a Blender-to-ComfyUI SDXL image-to-image pipeline, with usable parameter hints on denoise, ControlNet strength and model compatibility. Installation specifics are mostly deferred to the linked Notion document and GitHub pages.

20:01