In a previous post, I detailed the struggle of getting AI video models to respect the laws of chemistry. This week, an undergraduate attendee at my seminar on AI and video mentioned he was getting impressive results with Seedance 2.0. He was mostly testing it for “fun videos,” but I was still hopeful. I decided it was time to subject a new platform to my “molecule test.”
As is the case with most AI video companies, they offer very little in the way of free credits. I signed up for 800 credits for $20—which seemed like a fair deal for low-res testing. A 5-second test at 480p using the Seedance 1.5 model cost 20 credits.
The Version Confusion
I quickly realized a discrepancy: while the app itself is labeled v2.0, the 2.0 model is not actually available to me yet.
I ended up testing Seedance 1.5, and frankly, it was not great.
Even with a start and end frame provided—using the same prompt that gave me the best results in Lumalabs—the model just couldn’t maintain the structural integrity required for Organic Chemistry. An earlier prompt did the same thing with bunching up molecules and then kind of fading to the final structure. Part of the prompt said, “Move each molecule separately as it moves into final position” but it didn’t do that.
The Sora “Bonus”
The platform did offer a cool bonus: access to Sora. Having never been able to access Sora via OpenAI directly, I was eager to see if the hype translated to scientific accuracy. Interestingly, Sora doesn’t allow for an end frame, which was a drag for my specific workflow. I had to modify my prompt to focus on the movement toward the final positions.
The pricing for Sora was striking:
-
Sora 2: 10 credits for 10 seconds.
-
Sora 2 Pro: 80 credits for 10 seconds.
I opted for the Sora 2 Pro model, hoping the 8x price jump meant 8x the physics logic. After several hours of “They’re busy” messages, I finally got a result around 11:00 PM EST.
Looked nothing like what I expected but not having a starting frame made sense.
The Stress Test Prompt
I tried to be as explicit as possible to avoid the usual AI “hallucinations”:
“Have the molecules move to their final location and then have the connectors grow together. Please check the details for double bonds vs single bonds. Fade the text within 10 frames. Do not move the text. Do not modify the color of the molecules as they are moving. Just move the molecules into their proper positions. Move each molecule separately as it moves into final position – do not create extra molecules.”
The Verdict
The result? Horrible. Despite the “Pro” tag, the model couldn’t handle the separate movement of molecules or the specific bond types. It’s clear that for “fun videos,” these models are a blast, but for the bite-sized instructional content I’m building for Duke students, “different” hasn’t yet translated to “better.”
I tried an additional time starting with the first image and using a prompt generated by ChatGPT. I pasted the prompt at the end. The result was worse.
I’ll be waiting to see if Seedance 2.0 (the actual model, not just the app version) can finally provide the accuracy I’m looking for. Until then, the molecular mess continues.
Sora 2 Prompt Written By ChatGPT using the starting image:
Using the provided starting image of separated atom spheres (white = hydrogen, red = oxygen, black = carbon), create a smooth scientific animation where the atoms physically move through space and assemble into the correct molecular structure of acetylsalicylic acid (aspirin).
Requirements:
-
Do not morph or dissolve atoms. Each sphere must remain intact and travel naturally through 3D space.
-
Preserve atom colors and sizes:
-
White = Hydrogen (8)
-
Red = Oxygen (4)
-
Black = Carbon (9)
-
-
Begin with atoms arranged exactly as in the starting image (grouped by element).
-
Atoms should gradually translate, rotate, and bond into the accurate aspirin molecular geometry:
-
Benzene ring (6 carbons in hexagonal ring)
-
Carboxylic acid group attached to the ring
-
Ester (acetyl) group attached to the phenolic oxygen
-
-
End with a clean, stable ball-and-stick molecular model of acetylsalicylic acid centered in frame.
-
Camera: slow, subtle cinematic push-in (scientific visualization style).
-
Lighting: soft studio lighting, neutral gray background, realistic reflections.
-
Motion style: smooth, precise, educational, no jitter, no cartoon effects.
-
Duration: ~6–8 seconds.
-
Final frame should clearly show the complete aspirin molecule fully assembled and motionless.
Optional refinement (include if desired):
-
Very faint, subtle bond-forming glow when atoms connect.
-
Slight depth-of-field for professional molecular visualization look.
