11:42
Stable Diffusion 3.5
The video argues that Stable Diffusion 3.5 (Large and Large Turbo now, Medium on Oct 29) is usable today only in ComfyUI via a more complex triple-CLIP workflow that gives good-but-hit-or-miss image quality, while ComfyUI's new standalone .exe app brings real quality-of-life upgrades (verified updates, a bundled Manager, node version history, workflow-linked auto model downloads) even though it is still early beta.
SD3.5 ships in three sizes: Large, Large Turbo, and Medium (Medium releases October 29); the ComfyUI blog post covers the model links, clip model links, and an example workflow.
SD3.5 does not run on Forge (a normal KSampler is used with Model Sampling SD3, and the presenter states this combination does not work inside Forge UI).
Civitai's SD3.5 gallery lets you view prompts/settings per image and drag the image directly into ComfyUI to import its exact workflow; most Civitai images use the DPM++ 2M sampler with lower guidance scale and 40 steps, differing from the official ComfyUI workflow.
The SD3.5 ComfyUI workflow needs a triple CLIP loader including a new Clip G model (download separately into the clip folder); the other two clip models are reused from Flux setups.
The workflow includes a Conditioning Zero Out node combined with a Conditioning Set Timestep Range node whose purpose the presenter says he doesn't fully understand, especially since the negative prompt is empty.
The Empty Latent for SD3 node comes with a documented note: resolution should be a multiple of 64 pixels and total around 1 megapixel for good results.
The checkpoint has the VAE baked in; Model Sampling SD3 is set to shift=3 feeding a normal KSampler and a standard VAE Decode.
Image quality is described as pretty good but hit-or-miss across the Civitai gallery, often reading as a digital painting with 'a little bit of this kind of waxy skin'; a good render requires some luck.
Render times are surprisingly variable at identical resolution/pixel counts (43s vs 85s; 52s vs 76s on the presenter's RTX 4080), with no explanation offered.
ComfyUI.exe is a new self-contained, Electron-based standalone app no longer limited by browser constraints, giving the ComfyUI team full control over UI/UX.
New .exe features: verified/'code security' updates, a render history sidebar with image previews, a searchable/filterable node library, a model library with thumbnails, a saved-and-bookmarkable workflows panel, most-used-node-first search ordering, a floating pinnable queue/batch window, deep custom key-binding and color-palette settings, a toggle between the new floating 'top' menu and the classic menu, and a Manager now bundled directly into the app (with per-node-pack version history for rollback).
A workflow can be configured to link to where its models come from so the .exe auto-downloads them on first load — a timesaver — but this is opt-in per workflow and is not yet set up for the SD3.5 models shown, so those still require manual download.
Minor rough edges noted: the zoom-reset button seems of limited use next to mouse-wheel zoom, and toggled-off links render as straight lines rather than curves, making connections hard to trace; the presenter calls the app 'still in early beta.'
Stable Diffusion 3.5 Large — The largest of the newly released SD3.5 checkpoint sizes, with its VAE baked in, downloadable from the ComfyUI blog post or Civitai. Apply: Download the Large checkpoint plus its clip models (Civitai's download is noted as faster) and load it into ComfyUI's checkpoint loader.
Stable Diffusion 3.5 Large Turbo — A second SD3.5 size released alongside Large, positioned as a faster-generation variant. Apply: Swap the Turbo checkpoint into the same SD3.5 workflow used for Large when faster rendering is wanted.
Stable Diffusion 3.5 Medium — The third, not-yet-released SD3.5 size, scheduled for October 29. Apply: Not usable yet per the video; wait for the October 29 release before adding it to a workflow.
DPM++ 2M sampler — The sampler the video says most SD3.5 images on Civitai use, typically paired with a lower guidance scale and 40 steps. Apply: Select DPM++ 2M in the KSampler, lower the guidance/CFG value, and set steps to around 40 to match the results seen in Civitai's SD3.5 gallery.
Civitai drag-and-drop workflow import — A technique where a downloaded Civitai image embeds the full generation workflow, sampler, and settings used to create it. Apply: Download an image from Civitai's gallery and drag it into the open ComfyUI window to load its exact workflow.
Triple CLIP Loader with Clip G — The node in the official SD3.5 workflow that loads three CLIP models at once, including a new Clip G model not needed for Flux. Apply: Download the Clip G model into the ComfyUI clip folder and wire it into the triple loader alongside the two clip models already used for Flux.
Conditioning Zero Out + Conditioning Set Timestep Range — Two chained nodes in the SD3.5 workflow whose exact function the presenter says he isn't certain of, particularly since the negative prompt is empty. Apply: Include them as provided in the official/downloadable workflow rather than reconstructing manually, since their purpose isn't explained.
Empty Latent for SD3 — The SD3.5-specific latent-image node, documented as needing a resolution that is a multiple of 64 pixels and totals around 1 megapixel. Apply: Set width/height to 64-divisible values summing to roughly 1 megapixel of total pixels for good results.
Model Sampling SD3 (shift) — A node that configures SD3.5's sampling behavior, set to a shift value of 3 before feeding a standard/normal KSampler. Apply: Add the Model Sampling SD3 node ahead of the KSampler and set shift to 3, as shown in the video's workflow (noted not to work in Forge UI).
ComfyUI.exe (standalone Electron app) — ComfyUI's new self-contained desktop application, no longer running inside a browser. Apply: Install the .exe to access features unavailable in the browser build, such as unrestricted custom key bindings.
Verified updates / 'code security' — A .exe feature where every software update is verified before being applied. Apply: Accept updates inside the standalone app relying on this built-in verification, per the video.
Workflow-linked auto model download — A feature letting a workflow specify where its models come from so ComfyUI.exe downloads them automatically the first time the workflow is loaded. Apply: Add model source links inside a workflow file by hand to enable this; it is not yet configured for the SD3.5 workflow shown, so those models must currently be downloaded manually.
Queue/render history sidebar — A left-side panel listing previously rendered images with previews. Apply: Use it to reload a past image along with the generation settings that produced it.
Searchable/filterable Node Library — A panel of installable/available custom nodes inside the .exe, with search and filter controls. Apply: Search or filter here to locate nodes quickly while building a workflow.
Model Library with thumbnails — A panel listing installed models with thumbnail previews to aid identification. Apply: Browse by thumbnail rather than filename to find the right checkpoint or model faster.
Saved and bookmarked workflows panel — A section for saving workflows built in the app, with a separate bookmarks area for favorites. Apply: Bookmark frequently used workflows so they surface at the top for faster access.
Most-used-node search ordering — When double-clicking the canvas to search for a node, the app surfaces the most frequently used nodes first. Apply: Rely on this ordering to speed up repeated node insertion while building graphs.
Floating pinnable queue/batch window — A movable status window showing queue and batch number, with controls to cancel the current render, clear pending tasks, and refresh nodes/models; it can be docked to the top of the screen. Apply: Drag it anywhere convenient, or pin it to the top so it behaves like the classic top menu bar (toggle via the menu's 'top' setting).
Custom key bindings and color palette settings — A Settings section allowing full keyboard shortcut remapping and interface color palette changes, unrestricted by browser limitations. Apply: Open Settings in the .exe to rebind key combinations or change the color palette to preference.
Built-in ComfyUI Manager with node-pack version history — The node/package manager, now bundled directly into the .exe, including a view of every installed version number for a given custom node pack for rollback. Apply: Open Manager to install, list, or update nodes, and use the version history to roll back to an older node-pack version if a newer one breaks an existing workflow.
Toggle link visibility / select mode (space vs. pen) — Canvas controls to hide/show node-connection lines and to switch between space-bar-drag panning and a click-and-move 'pen' mode. Apply: Toggle link visibility off to declutter a crowded graph (noting links render as straight, not curved, lines), and use pen mode to pan/select without holding the space bar.
The Conditioning Zero Out + Conditioning Set Timestep Range combo ships in the official SD3.5 workflow but its function is unclear even to the presenter, especially with an empty negative prompt — suggesting parts of the reference workflow are copied/boilerplate rather than fully explained by ComfyUI or Stability AI.
Identical-resolution renders showing wildly different times (43s vs 85s, 52s vs 76s) point to an undiagnosed bottleneck in the SD3.5 ComfyUI pipeline; a commenter with an RTX 4090 independently reports the same unexplained long generation times, corroborating the pattern beyond the presenter's own hardware.
The .exe's flagship 'auto-download models from a workflow' feature is opt-in and must be configured inside each workflow file by hand — it doesn't retroactively make an already-shared workflow (like the SD3.5 one) self-installing, so the marquee convenience feature doesn't yet apply to the very models the video is testing.
A viewer comment fills a gap the video leaves open: most of the .exe's new UI, keybinding, and manager features are already usable in the browser version of ComfyUI, meaning the standalone app's real differentiator is being an Electron-based app rather than exclusive functionality.
«stabil Fusion 3.5 is out but is it any good»
— 00:00
«now here's an interesting note it says make sure the resolution is mul of 64 pixels and adds up to around 1 megapixel to get good results from this rendering»
— 03:25
«you can see that the vae is baked into the 3.5 model»
— 03:41
«this does not work inside of Forge UI»
— 03:58
«you will see that it can be a little bit hit and miss»
— 04:19
«it's surprisingly variable not quite sure why that happens»
— 04:54
«they have this kind of code security where any update is verified for you that already is a huge Improvement in that software»
— 06:05
«this is still in early beta and still has a way to go»
— 11:11
Reception
Audiences appreciate the channel and content delivery, but express significant skepticism about whether SD3.5 represents a meaningful improvement over existing models like Flux.
The video functions as a solid practical walkthrough of the SD3.5 ComfyUI workflow and a feature tour of ComfyUI.exe, but it stops short of explaining several nodes (Conditioning Zero Out/Timestep Range) and the render-time inconsistency it shows on screen, leaving exactly the questions its own audience raises in the comments unanswered.
11:42