I wanted to share a quick perspective on something that has drastically shifted my income over the last few months. Like a lot of people here, I’m always keeping an eye out for tech adjacent side income, but I stumbled into the AI data annotation space and the numbers have been honestly eye opening. I’ve been actively working across a few platforms specifically DataAnnotation, Snorkel, and Handshake. Coming into it, I was skeptical. We’ve all seen the low tier micro task sites that pay pennies. But because these specific platforms are training LLMs on complex reasoning, coding, and high level writing, the hourly rates for US-based freelancers are surprisingly substantial. The best part isn't just the money, though, it’s the absolute autonomy. I log on whenever I feel like it, work straight from my laptop, and pick the projects that actually interest me. Seeing how much corporations are pouring into AI development right now, it makes complete sense why they are paying a premium for quality human evaluation. It takes a lot of focus and sharp critical thinking to pass the initial assessments and keep getting high paying tasks, but the financial payoff has made it completely worth it. It’s wild to me that more people in the tech community aren't leveraging this right now. Curious if anyone else here is actively doing LLM training or prompt engineering evaluation? I'd love to compare notes on project availability, or fill in anyone who is trying to figure out how to break into the higher tier platforms.

A LoRA and a RefMod can both help with a character, but you get there differently A LoRA learns adapter weights through training. The Full Reference mode in this MiniMax H3 workflow encodes your images into a reusable reference file. No LoRA training run Loose analogy: a LoRA is practice, a RefMod is a reference sheet beside the drawing board. Neither stops the artist from occasionally getting things wrong I made RefMod Pilot to make that second route easier to try. Upload a few consistent photos, save the reference, then use it in a video workflow with Characters → Scene → Output. Scene settings include aspect ratio and duration in seconds. The technical nodes sit inside a subgraph There's a bundled orange robot example made from six views. Its reference took 5.7 seconds to encode on a 48 GB Blackwell MIG instance. That's one encoding measurement, not a video-generation benchmark. More references still add cost when generating, and the robot's finished-video likeness is untested I'm the developer of LoRA Pilot. RefMod Pilot is free to download, but you'll need the H3 models, dependencies and GPU compute. Credit for the underlying reference nodes goes to Luisacaotica GitHub: https://github.com/vavo/minimax-h3-refmod-pilot Hugging Face: https://huggingface.co/no3us/minimax-h3-refmod-pilot

Turn a still image into a video in your local ComfyUI, without a local inference GPU or downloading models. In this walkthrough, we install the free Forma custom nodes, connect an account, open an image-to-video template in Comfy UI and run a real generation. Your price depends on the model and settings - no subscription needed. Pay as you go - top-ups from $5. Check the quote before running the inference. Choose from 39 image and video models, 8 video workflow types and 2 image workflow types. Payment options include cards, Apple Pay, Google Pay, Link and crypto, depending on availability at checkout. Get the nodes, installation instructions and example workflows: https://github.com/vavo/runpod-endpoints-comfy

I've shared some of my LoRA Pilot work here. This is my other project, Forma, for the days you want to generate an image or animate a still without setting up a workflow. You choose a model, write a prompt and see the price before generating. Text to image, image editing and image to video are in the same studio. No subscription; you top up a balance when you need it. Your library is private by default, with no public feed. For a short ad, I'd start with a few stills, choose the composition, then animate that one. Saves paying for video attempts while you're still deciding what the shot should look like. I built it with occasional creators in mind. It's a generation tool, so you'll still want an editor for cuts, titles and sound. https://forma.cam

The first three Making R01 tutorials are live. We’re using our orange robot to work through LoRA Pilot (www.lorapilot.com), from the first session to preparing a dataset. 1. One robot, five worlds A first session in the LoRA Pilot workspace. https://youtu.be/_Z85_IgteLQ 2. Directing the scene Camera prompts in ComfyUI and a look at how closely the results follow them. https://youtu.be/Q3NH6ZSfbOM 3. Preparing the dataset Upload an archive, review the images and edit captions in TagPilot. This episode covers preparation; we don’t run training yet. https://youtu.be/07dfuzvv_Ik