10:43
ComfyUI
The video, from Olivio Sarikas, demonstrates ComfyUI's "Model Sampling Flux" node, which sits between the model/Lora and the KSampler and exposes Max Shift, Base Shift, Width, and Height controls; the video's claim is that tuning these values (independent of seed and prompt) gives you a lot more control over Flux, letting you fix or vary fine details and composition without a full re-roll, and that this is especially useful with the 8-step Turbo Lora workflow.
Model Sampling Flux is a ComfyUI node placed in between the model and the KSampler; it can go before or after the Lora with no real difference, and it influences how the model renders the image.
The node's Width and Height values are explicitly stated to NOT be the resolution of the output/latent image; per Olivio they affect how the model renders (possibly something like an internal base image it references), though he says he is 'not 100% sure' how it works.
Testing the node's resolution setting (512, 768, 1024, 1600, 2048) with the same seed, settings, and prompt changes fine details and sometimes the overall composition.
At higher resolutions in this node (e.g., 2048x2048) there is a smaller usable range between Max Shift and Base Shift before the image quality degrades ('things go astray').
Max Shift and Base Shift are demonstrated on three example prompts (a skull/crown woman, a fish in a glass cup, and an anime fox thief) showing that shifting values can fix specific details (e.g., a better-looking fish, a tail correctly positioned behind clothing instead of poking through it, a more 'sinister' facial expression) while composition and colors largely stay the same.
The technique is framed as especially useful for the 8-step Turbo Lora workflow, giving more room to fix or improve results despite the compressed step count.
Olivio provides a free basic workflow (Google Drive) and a Patreon-exclusive batch workflow that auto-labels test images with their parameter values.
The batch workflow converts Max Shift and Base Shift into node inputs, uses a float-range node to step Max Shift from 0.75 to 1.75 in 0.1 increments (Base Shift held at 0.5), converts the floats to strings, and overlays the resulting text onto each output image by extending the canvas 50px at the top.
There is no single ideal setting according to the video, but Olivio states that higher node resolution narrows the usable Max/Base Shift range, and that values can also be inverted (lower Max Shift than Base Shift) for different results.
Olivio's personal preference, stated at the end, is a resolution of 1024x1024 in the node because it gives 'a nice big range' to play with Max/Base Shift values.
Model Sampling Flux (node) — A ComfyUI node inserted between the model/Lora and the KSampler that changes how Flux renders an image via Max Shift, Base Shift, Width, and Height settings, without altering the seed or prompt. Apply: Place the node in the workflow chain between the checkpoint/Lora loader and the KSampler, then tune its Max Shift, Base Shift, Width, and Height values before sampling to fix or vary details.
Max Shift — One of two paired numeric values in the Model Sampling Flux node that, when changed, alters composition and fine detail in the rendered output. Apply: Sweep Max Shift across a range (e.g., 0.75 to 1.75) while holding Base Shift, seed, and prompt constant to find a value that fixes an unwanted detail without discarding the overall composition.
Base Shift — The second numeric value in the Model Sampling Flux node, used alongside Max Shift to influence rendering; the video mostly holds it constant (e.g., 0.5) while testing Max Shift. Apply: Fix Base Shift as a baseline while varying Max Shift, or invert the pair (a lower Max Shift than Base Shift) to explore different composition/detail outcomes.
Width/Height (Model Sampling Flux node inputs) — Two settings in the node that, per Olivio, are not the output/latent image resolution but instead affect how the model renders the image, and that interact with the usable Max/Base Shift range — higher values narrow that range. Apply: Test different node Width/Height values (512, 768, 1024, 1600, 2048) with a fixed seed and prompt to observe detail changes, and default to 1024x1024 for the widest usable range when experimenting with Max/Base Shift.
Turbo Lora / 8-step Flux workflow — A Lora that lets Flux generate in as few as 8 steps ('the turbo model with eight steps'), used as the base setup the video tests Model Sampling Flux against. Apply: Load the Turbo Lora with the Flux model and use Model Sampling Flux's Max/Base Shift to compensate for or fix detail issues introduced by the compressed 8-step generation.
Batch parameter-sweep workflow (float range + string conversion + text overlay) — Olivio's Patreon-exclusive workflow that automates testing by converting Max Shift/Base Shift into external inputs, stepping Max Shift via a float-range node (0.75 to 1.75 in 0.1 increments), converting the floats to strings, and burning the resulting label into each output image by extending the canvas 50px and adding text. Apply: Expose Max Shift and Base Shift as node inputs, drive one with a float-range node across a step sequence, then use string-conversion and text-overlay nodes to label each generated image with its parameter values so a full batch can be compared visually.
The node's Width/Height inputs are decoupled from actual output resolution and instead function as a sampling/sigma-related control — a distinction confusing enough that it became a point of debate in the audience comments over what it's actually optimizing.
Max Shift/Base Shift function less like a full re-roll and more like a targeted 'detail fixer' dial: in the demonstrated examples, the overall composition and color palette stay consistent while specific problem details (an odd fish shape, a tail clipping through clothing) get corrected.
The relationship between resolution and shift range is inverse and non-obvious: raising the node's resolution setting shrinks the safe range of Max/Base Shift values before outputs break down, meaning the same numeric shift range behaves differently depending on the resolution chosen.
The batch-testing setup (float range → string conversion → text overlay baked into the image) turns an otherwise opaque numeric parameter search into a reproducible, self-labeling visual grid, directly addressing the kind of 'I don't know what the numbers do' confusion visible in the audience reaction.
«this will give you a lot more control over flux»
— 00:00
«something that is lit AF as they say»
— 00:06
«so this is of course in between the model so you can set it before the Lura after the Laura doesn't really make any difference it influences how the model is rendering your image»
— 00:32
«so this width and height is not the resolution of the latent image»
— 00:52
«but it can help you to fix problems in your image and have more control while having a consistent composition»
— 01:12
«this is extremely useful especially with the eight step Laura this can really enable you to get a much better image out of Flux»
— 06:57
«one thing I figured out is that when you have a higher resolution for example 2048 by 2048 you have a smaller range between Max and base that you can use before things go astray»
— 09:40
«you can also invert that so for example you can have a lower max value than the base value that sounds a little bit strange but experiment with that you can get some really interesting results from that»
— 09:54
«personally I found for myself that the value of 1024 by 1024 works best and gives you a nice big range on how to play with these values»
— 10:09
Reception
Technical community appreciates the parameter testing and detailed explanations, though some find the topic confusing or wish for deeper insights.
The video is a visual, trial-and-error walkthrough of ComfyUI's Model Sampling Flux node rather than a technical explainer of its mechanism — Olivio shows before/after comparisons across Width/Height and Max/Base Shift values and shares a batch-testing workflow, but he repeatedly flags his own uncertainty ('I'm not 100% sure') and states plainly that there is no single ideal setting, leaving the practical takeaway as 'experiment within these ranges' rather than a settled rule.
10:43