19:49
AI image generation
Matt Wolfe demonstrates a much faster, cheaper way to train a personal LoRA on the Flux AI image model via Replicate.com, letting him generate realistic AI images of himself (including composited scenes and animated video via Runway Gen 3) in about 26 minutes for roughly $2, replacing the 2+ hour Dream Booth/Stable Diffusion 1.4 process he'd used a year earlier.
Flux is, in the video's assessment, one of the best current models for photorealism, on par with Midjourney.
A year prior, training a personal likeness used Dream Booth on Stable Diffusion 1.4 inside Google Colab, took 2+ hours, and required keeping the browser open/active or the trained weights would disappear on timeout.
The new method trains a Flux LoRA on Replicate.com using Luca Taco's 'AI toolkit for flux Lora training' model, renting an Nvidia A100 GPU at roughly $5/hour ($0.001/sec).
Training requires a minimum of 12 images (Wolfe used 20), a Hugging Face account with a fine-grain access token, and an HF_REPO_ID setting to auto-upload the trained LoRA to Hugging Face.
The Hugging Face model repo must be set to public visibility, or Replicate cannot access it.
Actual training took 26 minutes and cost about $2.18; Wolfe offered viewers a $10 Replicate coupon code to cover this and further usage.
Generating an image afterward costs about 9 cents; default generation settings include 28 inference steps, a 1.0 LoRA scale, 1-4 output images, and configurable aspect ratio (e.g., 16:9).
Placing the trained trigger word as the very first word of a prompt produces noticeably better/more reliable results than placing it elsewhere in the prompt.
Wolfe used a Claude 'projects' setup with persistent custom instructions to help write/optimize prompts.
A finished Flux LoRA image can be animated by taking the same prompt into Runway Gen 3 and setting the image as the first frame, producing a short video (e.g., Wolfe and Deadpool walking from an explosion).
Flux — Flux is a text-to-image AI model that the video says is 'one of the best models at generating realism that exists right now,' comparable to Midjourney. Apply: Use Flux as the base model to generate photorealistic images, then fine-tune it with a personal LoRA to insert your own likeness.
LoRA (Low-Rank Adaptation) training — LoRA is the fine-tuning method used to train a small add-on model that teaches Flux a person's face/likeness without retraining the full base model. Apply: Collect a set of your own photos (the video used 20, states 12 minimum) and run them through a LoRA training model to produce a personal Flux LoRA.
Replicate.com — Replicate.com is a platform where users can rent GPU compute (e.g., an Nvidia A100 at roughly $5/hour) to run AI training and inference jobs. Apply: Sign up on Replicate, use it to both run the Flux LoRA training job and later generate images from the trained model, paying per-second of GPU use.
AI toolkit for flux Lora training (Luca Taco's Replicate model) — A specific pre-built Replicate model, created by Replicate contributor Luca Taco, packaged for training Flux LoRAs from a user's image set. Apply: Select this model on Replicate, upload your training images, and configure settings like training steps and repo ID to launch training.
Hugging Face fine-grain access tokens — A Hugging Face account feature allowing creation of scoped API tokens with selective (fine-grained) permissions instead of full account access. Apply: Create a fine-grain token on Hugging Face and supply it to the Replicate training job so it can upload/access your model repo.
HF_REPO_ID auto-upload setting — A configuration field in the training job that, when set to a Hugging Face repo ID, automatically uploads the resulting trained LoRA to that Hugging Face repository. Apply: Set HF_REPO_ID before starting training so the finished LoRA lands directly in your Hugging Face account, and ensure repo visibility is public so Replicate can access it.
Trigger-word-first prompt placement — An informally observed technique, self-described by Wolfe as possibly coincidental, where placing the model's trigger word as the very first word of a prompt produces more reliable inclusion of the trained likeness than placing it elsewhere. Apply: When writing prompts for your trained model, lead with the trigger word (e.g., 'Mr eow and Deadpool...') rather than embedding it mid-sentence.
Claude Projects (custom instructions) — A Claude feature that lets a user set persistent custom instructions for a project, used here to help optimize/generate image prompts. Apply: Set up a Claude project with instructions tailored to writing Flux prompts, then use it to draft or refine prompts before generating images.
Runway Gen 3 — A video generation tool that can animate a static image into a short video clip. Apply: Take a finished Flux LoRA image, reuse its original prompt, and feed the image into Runway Gen 3, set as the first frame, to generate an animated video.
Dream Booth — The older training method, used with Stable Diffusion 1.4 in the video's prior tutorial, that required 2+ hours in a Google Colab notebook that had to be kept open to avoid timeout. Apply: Cited in the video only as the outdated baseline being replaced by the new Replicate/Flux LoRA workflow, not recommended for current use.
Google Colab — A hosted Jupyter notebook environment previously used to run the Dream Booth training process, with a Pro plan available for $10/month for A100 access. Apply: Referenced as the old way to access an A100 GPU for training; the video steers viewers to Replicate instead for the current method.
The trigger-word-first placement technique is explicitly hedged by Wolfe himself as possibly coincidental ('I don't know if this is how this actually is supposed to work... that could be totally coincidental'), making it an empirically-observed tip rather than documented behavior.
The video cites internally inconsistent cost figures for training — an actual paid cost of $2.18, alongside a separately mentioned '$225 estimated training cost' figure — without reconciling the two, which is worth noting as stated rather than resolved.
The full workflow chains four distinct platforms/tools, each doing one narrow job: Hugging Face (storage/hosting), Replicate (training + inference compute), Claude projects (prompt engineering assistance), and Runway Gen 3 (animation) — a modular pipeline rather than a single all-in-one tool.
Wolfe frames the improvement primarily in terms of attention cost, not just money or time: the old method required actively babysitting a browser tab to prevent timeout, while the new method runs unattended on rented cloud compute.
«this is super cool I figured out a way to use the new flux AI image generation model and actually train myself into the model so that I can get pictures like this of me and Deadpool walking away from an explosion»
— 00:00
«in my opinion flux is one of the best models at generating realism that exists right now it's pretty on par with what we get out of mid journey»
— 01:03
«the process to do that took a good two plus hours you had to train it on dream Booth inside of a Google collab you had to keep the browser open and kind of scroll the website every once in a while to make sure the website didn't time out»
— 02:16
«it cost about a tenth of a penny per second or roughly $5 an hour however I'm going to hook you up with a coupon code»
— 04:36
«when I put Mr eow my trigger word as the first word it works a lot better than if it's somewhere else within the prompt»
— 17:52
«flux is fun trending your own likeness into the flux images is even more fun and then turning it all into a video uh it it's kind of awesome»
— 19:07
Reception
Audience appreciates the tutorial content but is frustrated by the repeatedly expiring free credit link that breaks the core value proposition of the video.
The video is a hands-on product tutorial chaining several paid platforms (Replicate, Hugging Face, Runway) into a step-by-step method for training a personal Flux LoRA and animating the results, offered as free/cheap via a limited-time coupon rather than a fully free method in itself.
19:49