Mark Belden (@markbnyc573) profile picture

Creative Engineer @trymelius

NYC
Joined September 2025
Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)Photo by Mark Belden (@markbnyc573)
Messing around with blender to see how it would improve text and product accuracy in video generations. Adds a step to the workflow, but definitely worth it if you are having trouble with garbled text or product consistency.
🤖 Made with AI
3
7
214
I've spent at least 5m in AI image/video gen credits this year. Here's everything I've learned so far A thread ↓
13
13
352
You don't need the top models to make it work, though. I've seen customers make great AI UGC micro-dramas with just Seedance 2.0 and Nano Banana as their workhorses.
3
27
8. 90% of my video gens are Seedance 2.5, because I'm spoiled and have unlimited credits. On a budget, I'd save it for the shots that matter: product consistency and complex movement. It's still worth the price for the output.
2
31
7. Pick your image model by the job. - Text matters (static ad units, packaging): GPT Image 2.5 - People, realism, and fantasy: Seedream 5.0 I barely use Nano Banana Pro anymore. I used to love it. Patiently waiting for the update.
2
37
6. Blender is great for building references. Block out your scene and camera there, then use the renders to guide your video model. There are tons of tutorials on this. Look up Blender + GPT-6 Astra + Seedance 2.5 if you're curious.
2
26
5. Video to video REFERENCES are good. (the one v2v I'll defend): put your voiceover on a black video and run it as video-to-video. The output lines up with the real VO, no offset. Build pauses into the audio, not the prompt. If the prompt asks for a pause, it re-voices the line.
2
12
4. Native AI voices mangle brand names. Seedance said "Mel-us" instead of "Mel-e-Us." Use ElevenLabs with a phonetic spelling ("Mel-e-us").
2
16
Even a character holding the product up near their face can garble the label. The reliable fix is in post: motion-track the real artwork back on (After Effects, Mocha).
2
19
3. AI video still can't do fine text on logos and packaging. - Start with the highest-res, front-facing product shot you have - Keep the pack still and large in frame - One run on a premium model beats rerolling a cheap one 3x
2
23
Before you reroll, ask yourself: can I save this by editing it myself? If you do reroll, steer it with images and text, not text alone. Trim to the last good frame, fix that frame as a still, regenerate only from there, then stitch.
2
28
2. Video-to-video edits are almost always a no. I review a lot of customer canvases and see tons of them. AI is non-deterministic, so you end up playing whack-a-mole: fix one issue, introduce two more. Or make the original worse.
2
28
1. Animatics still matter. Pull your stills and audio and have Claude or ChatGPT assemble a rough cut. Play it back. If there are gaps, fix them now. Don't count on the AI to get it right. You get client sign-off without burning 40k credits on footage that gets cut.
2
43
Kling 4 vs Seedance 2.5
6
93
we have some real talent at the office.
3
11
243
Here's how I handle incorrect product sizes in with AI image models.
3
9
185
Mark Belden retweeted
Krea 3, the new SOTA image editing model, is now on Melius. Key winning points: - Supports bounding boxes and regional edits - Capable of complex edits with 6 reference images - GENERATIONS IN ~6 SECONDS This model is pushing the frontier. Try it today.
9
3
58
1,835
Mark Belden retweeted
every startup settings page is a museum of customers who complained once
8
2
56
1,631
I make videos by talking to my computer now. I don't drag and drop. I don't click. I just let my mind flow. Throughout my editing career there has always been a moment where I thought, "I wish I had this shot." Now, I can actually have that shot. Or any shot. Or any animation. I've never felt so empowered to execute my vision from start to finish.
2
4
123
Mark Belden retweeted
Cast is an incredible milestone for us. We've reached SoTA on realistic camera work, character consistency, script adherence, and world building using our agent and the Cast interface. Incredibly proud of the team! Look at this microdrama episode we one-shotted :)
Introducing Cast, a platform for anyone to make anything with AI Avatars AI UGC, podcasts, microdramas, all accessible through an easy interface where you just fill in the blanks. Up to $50 in free credits if you sign up for a paid plan. Available now, on Melius.
5
4
57
3,748