20:58
ComfyUI
The video is a step-by-step tutorial for setting up ComfyUI locally so a user can generate images and videos entirely on their own computer, offline, for free, using largely uncensored open-weight models like Flux and the Wan/Yi 2.2 family — avoiding data sharing with big tech companies and any generation costs or subscriptions.
Running image/video generation locally on your own computer means no data goes to Google, Microsoft, OpenAI, or other big tech companies, and they cannot train on anything generated this way.
Local generation is useful for remote or no-internet situations (e.g. generating images at 30,000 ft on a flight) and for private use cases.
Comfy UI is a visual, node-based control panel for AI image/video generation that is 100% free and open source forever, with no subscriptions or hidden costs.
Comfy UI workflows look like an overwhelming 'spaghetti bowl' of connected nodes, but the video deliberately sticks to the simplest possible workflows and pre-built templates so understanding the node architecture isn't required.
Flux 1 Craya Dev is presented as the best image generation model currently runnable locally.
The Flux family of models is largely uncensored but does not generate accurate likenesses of celebrities.
Mac and PC require different model files due to architecture differences; FP8/Float8 versions need CUDA (Nvidia) and are incompatible with M-series Macs.
Video generation on Mac hardware is very slow — 20 to 60 minutes for a single 6-second clip.
The Yi 1.2.2 family is described as the most powerful open-weight video model family currently available, with 14B and 5B parameter variants suited to different GPU capabilities.
Smaller 5B video models generate faster at a slight cost to quality compared to 14B models.
Generated outputs remain local on the user's hard drive unless the user explicitly enables cloud sync.
Open-weight models are being fine-tuned by the community to produce content that would otherwise be censored or restricted.
The video positions itself as a deliberately simple, fast on-ramp to Comfy UI rather than a deep technical tutorial.
Using Comfy UI reportedly puts a user ahead of 99% of people who only know consumer tools like ChatGPT and Veo.
Deeper technical mastery of Comfy UI (additional nodes, optimization techniques) is available via other named creators' channels.
Wolfe voices personal unease about how far this technology could be taken, while still choosing to teach it, framing it as inevitable ('you can't put this genie back in the bottle').
Comfy UI — A visual, node-based control panel for creating AI-generated images and videos, described as 100% free and open source forever with no subscriptions. Apply: Install it from comfy.org (the simplest route) or via a manual GitHub install, then use pre-built templates to generate images/videos without needing to understand the underlying node graph.
Flux 1 Craya Dev — The image generation model the video calls the best one currently runnable locally. Apply: Download its ~23.8 GB diffusion model (.safetensors) file and load it into a Comfy UI template to generate images offline.
Flux models (family) — A family of largely uncensored image generation models that do not accurately render celebrity likenesses. Apply: Use a Flux-based Comfy UI template for general uncensored local image generation, but don't expect accurate real-person likenesses.
Stable Diffusion models — An alternative image generation model architecture mentioned alongside Flux as another Comfy UI option. Apply: Select a Stable Diffusion-based template in Comfy UI as an alternative to Flux for local image generation.
Yi 1.2.2 14b text-to-video — A 14-billion-parameter video model variant that generates video directly from a text prompt. Apply: Use this variant when you have a robust GPU and want to create video purely from a written prompt.
Yi 1.2.2 4b image-to-video — A video model variant that animates a static image into a video clip. Apply: Feed a starting image into this model within Comfy UI to turn it into a short video.
Yi 1.2.2 5b video generation — A 5-billion-parameter video model combining text-to-video and image-to-video capability, suited to older or slower Nvidia GPUs, generating faster with a slight quality reduction versus 14B variants. Apply: Choose the 5b variant when your GPU is weaker and you want faster generation at a modest quality tradeoff.
Yi 1.2.2 14b first-and-last-frame-to-video — A video model variant that interpolates motion between two supplied key frames (a first and last frame). Apply: Supply a starting image and an ending image to this model to generate the video connecting them.
Mochi — A video generation model listed among the options available for AI video creation. Apply: Consider Mochi as an alternative video model when exploring Comfy UI's video generation templates.
Hunion video — A video generation model listed among the available Comfy UI options (transcribed as 'Hunion video'). Apply: Use it as an alternative to Wan/Yi models within Comfy UI video templates.
V3 model — A video generation alternative noted to produce better-quality results on certain specific prompts. Apply: Try the V3 model when the default Yi/Wan model output underperforms on a particular prompt.
Halo AI model — Another video generation alternative mentioned alongside the V3 model for prompt-specific quality comparisons. Apply: Test the Halo AI model as a comparison point against Yi/Wan and V3 for a given video prompt.
Recraft models — An API-based (cloud, non-local) image generation model option. Apply: Use Recraft via an API connection in Comfy UI when local generation isn't needed or preferred.
Runway models — An API-based video generation model option. Apply: Connect to Runway via API in Comfy UI for cloud-based video generation instead of local compute.
Stability models — An API-based generation model option from Stability AI. Apply: Use Stability's API connection in Comfy UI as a cloud alternative to local models.
Idiogram — An API-based image generation model option (Ideogram, transcribed as 'Idiogram'). Apply: Connect to Ideogram via API within Comfy UI for cloud image generation.
MPS backend — Metal Performance Shaders, the GPU acceleration backend Comfy UI uses on Mac hardware instead of CUDA. Apply: Rely on MPS automatically when running Comfy UI on an M-series Mac; don't select CUDA/FP8-only model variants.
Safe tensors — The model file format (.safetensors) used to distribute diffusion model weights, e.g. the ~23.8 GB Flux 1 Craya dev.safetensors file. Apply: Download the correct .safetensors file for your chosen model and place it in the appropriate Comfy UI models folder.
VAE — Variational Autoencoder, a model component used in the image/video generation pipeline. Apply: Download and load the matching VAE file alongside the diffusion model in the Comfy UI workflow.
Text encoders — The language-model component responsible for interpreting text prompts within the generation pipeline. Apply: Download and load the text encoder file required by your chosen diffusion model template.
Diffusion models — The core generative architecture underlying image/video models like Flux, distributed as a large model file (e.g. 11 GB). Apply: Place the diffusion model file in Comfy UI's models directory so templates can reference it.
CUDA architecture — Nvidia's GPU computing architecture, required to run FP8/Float8 model versions. Apply: Only choose FP8 model variants when running Comfy UI on an Nvidia GPU; Mac users must select different, non-CUDA versions.
Hugging Face — An online model repository used as a source for downloading diffusion models, VAEs, and text encoders. Apply: Download model files directly from Hugging Face pages linked in Comfy UI template instructions.
CivitAI — An online model repository (transcribed in the source as 'Civot AI') for finding additional models. Apply: Browse CivitAI as an alternative source for community-shared models and fine-tunes.
GitHub manual install — A more technical alternative method for installing Comfy UI via its GitHub repository instead of the simple comfy.org installer. Apply: Use the GitHub manual install route if you need more control than the standard installer provides.
Workflows — Node-graph templates within Comfy UI that define a full generation pipeline, which can look like a 'giant spaghetti bowl of connected nodes.'. Apply: Load a pre-built workflow rather than constructing node graphs from scratch to start generating quickly.
Templates — Pre-configured workflows bundled with Comfy UI that remove the need to understand the underlying node architecture. Apply: Select the built-in template matching your goal (e.g. Flux image generation or Yi/Wan video generation) as your starting point.
API connections — Cloud-based model calls (e.g. to Recraft, Runway, Stability, Ideogram) usable inside Comfy UI as an alternative to local generation. Apply: Configure an API connection node in Comfy UI when you want to use a cloud model instead of a locally downloaded one.
Negative prompts — Instructions within a generation prompt specifying what should NOT appear in the output. Apply: Add a negative prompt alongside the main prompt to steer the model away from unwanted elements.
Comfy UI nodes — The composable building blocks that make up a Comfy UI workflow graph. Apply: Learn additional node types (via recommended deeper-dive channels) to customize workflows beyond the provided templates.
Comfy UI optimization strategies — Techniques for improving Comfy UI performance and workflow efficiency, referenced as a deeper topic beyond this video's scope. Apply: Consult the recommended AI Entrepreneur or Olivio Sarais channels for concrete optimization techniques not covered in this tutorial.
The privacy pitch is framed as a specific technical claim, not generic privacy talk: because nothing is sent to the companies, they literally cannot train on anything generated this way.
Hardware compatibility functions as a hidden gatekeeper in the tutorial: model choice isn't just about quality but about a hard Mac/PC (CUDA vs. MPS) fork that determines which file variant will even run.
Model capability and file size scale together in the video's framing (an 11 GB diffusion model vs. a ~23.8 GB Flux file), implying users trade download/storage cost for output quality.
The tutorial markets offline/uncensored capability as a selling point while, near the very end, the same creator flags that this same lack of restriction is what worries him about where the technology is headed — the sales pitch and the caveat coexist rather than the caveat gating the pitch.
Wolfe explicitly outsources 'going deeper' (advanced nodes, optimization) to other creators (AI Entrepreneur, Olivio Sarais), positioning his own video as intentionally shallow and fast rather than comprehensive.
«Once you have this set up, you won't even need the internet to generate images and videos. It all happens completely on your computer.»
— 00:10
«The models we'll be playing with are pretty much completely uncensored, so you can generate any weird abomination that pops into your head.»
— 00:17
«Comfy UI is 100% free and open source and always will be. No subscriptions, no hidden costs.»
— 02:18
«The absolute best image generation model that can be run locally on your computer is this Flux 1 Cade Dev.»
— 05:25
«I do have a little bit of concern about where this is all headed. I mean, you could literally generate anything you want.»
— 18:23
«You can't put this genie back in the bottle.»
— 18:50
«if you start playing with Comfy UI, you're going to be ahead of 99% of the rest of the world who basically only knows how to use tools like ChatgPT and VO»
— 19:20
«this was meant to be the simplest overview, the fastest way to get up and running with Comfy UI, generating images and videos offline»
— 20:09
Reception
Audience is interested in the tool but overwhelmingly frustrated by installation failures, platform incompatibility, and undisclosed costs.
This is a practical walkthrough of a genuinely free, open-source local pipeline (ComfyUI plus Flux and Wan/Yi models) for offline, largely uncensored image and video generation; the video's own closing caveat about misuse risk is presented as the creator's stated concern rather than a gate on the instructions themselves.
20:58