Here's a workflow you've probably wanted: describe a scene, get a polished still image back, then animate it into motion — all in one Claude Code session, without opening a separate tool.
That's image-to-video for coding agents. The still image becomes the first frame. The video model animates it. Your agent handles both steps.
But the pipeline isn't just two commands chained together. The model pairing matters. Seedream 5 generates differently than Nano Banana Pro. Veo 3.1 animates differently than Kling 3.0. Getting the combination right makes the difference between a clip that looks like a demo and one that looks like a draft.
This guide covers the whole pipeline: which image models pair best with which video models, when to use text-to-video instead, and how to run the entire workflow in one agent session.
Why Image-to-Video Beats Text-to-Video Alone
Text-to-video sounds simpler. One prompt, one clip, done. And for quick social content or conceptual previews, it works.
But text-to-video gives you less control. You describe a scene. The model interprets it. If the interpretation is off — if the composition is wrong, the lighting doesn't match, the subject position feels awkward — you start over with a different prompt and hope for a better roll of the dice.
Image-to-video separates the two concerns:
The still image defines the composition. You generate a keyframe. You check it. If the composition is wrong, you regenerate just the image — not the whole video.
The video model adds motion. Once the still looks right, you feed it to the video model. The motion can be subtle (a slow push-in) or dramatic (a tracking shot through a scene). Either way, the starting frame is locked.
This two-step workflow gives you editorial control. You approve the frame before committing motion budget to it. For anything that matters — product demos, landing page hero clips, pitch deck visuals — that control is worth the extra step.
The Pipeline: Step by Step
Step 1: Choose Your Still Image Model
You have seven image models available through AnyCap. For image-to-video workflows, three stand out:
| Model | Why for image-to-video | Best use |
|---|---|---|
| Seedream 5 | Strongest first-pass quality. The still looks closer to final with less iteration. | When the keyframe will be the foundation of a customer-facing video. |
| Nano Banana Pro | Best for revision loops. Generate, evaluate, tweak, repeat — the edit workflow is smoother. | When you're iterating on a concept and want to try variations before animating. |
| Nano Banana 2 | Fastest generation speed. Less polish per image, but you can try more compositions in the same time budget. | When you're exploring concepts and want volume over perfection. |
The rule of thumb: if the video is customer-facing (demo, announcement, teaser), start with Seedream 5. If you're exploring or prototyping, start with Nano Banana 2 and upgrade the winner.
Step 2: Lock the Keyframe
Generate the still. Evaluate it. Don't proceed to video until the composition, lighting, and subject position are correct. Here's a practical workflow:
# Generate three keyframe options with different compositions
anycap image generate \
--prompt "a modern SaaS dashboard on a laptop, floating UI elements, clean studio lighting, product photography style" \
--model seedream-5 \
-o keyframe-1.jpg
anycap image generate \
--prompt "same dashboard, angled perspective from above, softer lighting, more depth of field" \
--model seedream-5 \
-o keyframe-2.jpg
anycap image generate \
--prompt "same dashboard, dark mode, neon accent colors, dramatic side lighting" \
--model nano-banana-2 \
-o keyframe-3.jpg
Review all three. Pick the best one. Now you have a locked keyframe.
Step 3: Choose Your Video Model
Different video models handle image-to-video differently. The source image matters as much as the motion style you want:
| Video Model | Image-to-Video Style | Best Pairing |
|---|---|---|
| Veo 3.1 | Smooth, polished motion. Handles subtle camera moves well. | Seedream 5 — premium still → premium motion |
| Seedance 1.5 Pro | Steady, production-repeatable. Reliable frame-to-motion translation. | Nano Banana Pro — consistent revision → consistent motion |
| Seedance 2.0 | Newer model, stronger cinematic feel. Better at interpreting depth in the source still. | Seedream 5 or FLUX.1 Kontext Max |
| Kling 3.0 | Strongest camera dynamics. Controllable pan, zoom, and tracking. | FLUX.1 Kontext Max — rich still → dramatic motion |
| Kling O1 | Image-first design. The source frame drives the entire video. Good for product shots. | Nano Banana Pro or Seedream 5 |
| Sora 2 Pro | OpenAI's best. Handles complex scenes and realistic motion. | Seedream 5 — maximum quality pipeline |
Step 4: Animate
Feed the keyframe to the video model with a motion prompt:
anycap video generate \
--prompt "slow push-in toward the laptop screen, UI elements animate sequentially, smooth parallax on background" \
--model veo-3.1 \
--mode image-to-video \
--param images=./keyframe-1.jpg \
-o demo-clip.mp4
The prompt describes motion only — not the scene. The scene is already locked in the keyframe. Describe what the camera does, how elements move, what changes over time.
Model Pairing Matrix: Which Image + Which Video?
Here's the full pairing grid. Each combination has a different feel and fits a different workflow:
| Veo 3.1 | Seedance 2.0 | Seedance 1.5 Pro | Kling 3.0 | Sora 2 Pro | |
|---|---|---|---|---|---|
| Seedream 5 | ⭐ Premium pipeline. Best possible output. | Strong cinematic feel. Good for brand videos. | Reliable, slightly less motion flair. | Dramatic motion from polished stills. | Maximum quality, highest cost. |
| Nano Banana Pro | Clean motion from edited stills. | Good for iterative revision → motion loops. | ⭐ Best revision-to-motion workflow. | Bold motion treatment of refined images. | Solid, if you prefer the OpenAI stack. |
| Nano Banana 2 | Fast iteration → decent motion. | Quick draft pipeline. | ⭐ Best for prototyping at speed. | Dramatic drafts from rough stills. | Overkill for draft-quality stills. |
| FLUX.1 Kontext Max | Rich visual → polished motion. | Design-heavy motion. | Steady treatment of rich visuals. | ⭐ Best cinematic pipeline. | Premium design-to-motion. |
| GPT Image 2 | Solid if you prefer OpenAI stack. | Good if both models are OpenAI-preferred. | Reliable cross-stack output. | Interesting crossover. | ⭐ Full OpenAI pipeline. |
⭐ = recommended pairing for that workflow type
Three Real Pipelines, End to End
Pipeline 1: Product Demo Clip (Customer-Facing)
Goal: Generate a polished product demo video for a launch page.
# Step 1: Generate the hero keyframe
anycap image generate \
--prompt "product shot of a web application dashboard on a MacBook, floating data visualizations, clean modern office background, soft natural light, product photography" \
--model seedream-5 \
-o hero-frame.jpg
# Step 2: Animate with subtle camera move
anycap video generate \
--prompt "slow gentle push-in toward the screen, data points appear one by one, subtle parallax on the background window" \
--model veo-3.1 \
--mode image-to-video \
--param images=./hero-frame.jpg \
-o product-demo.mp4
# Step 3: Store and share
anycap drive upload product-demo.mp4
Result: A 10-second clip with the production quality of a commissioned video — generated in one session. The still image locked the composition. Veo 3.1 added smooth, polished motion.
Why this pairing: Seedream 5 gives you the strongest still. Veo 3.1 gives you the smoothest motion. Together, they produce output that looks professional even before post-production.
Pipeline 2: Social Content Batch (Volume)
Goal: Generate 10 short-form video variants for A/B testing on social.
# Step 1: Define a batch prompt template
PROMPT_BASE="bold social media announcement graphic, vibrant colors, clean typography area, modern design style"
# Step 2: Generate 3 keyframe variants (fast)
for i in 1 2 3; do
anycap image generate \
--prompt "${PROMPT_BASE}, variant ${i}" \
--model nano-banana-2 \
-o social-frame-${i}.jpg
done
# Step 3: Animate each variant with different motion
for i in 1 2 3; do
# Version A: subtle zoom
anycap video generate \
--prompt "gentle zoom-in, text elements fade in" \
--model seedance-2.0-fast \
--mode image-to-video \
--param images=./social-frame-${i}.jpg \
-o social-${i}a.mp4
# Version B: pan across
anycap video generate \
--prompt "slow pan left to right, elements slide in from edges" \
--model seedance-2.0-fast \
--mode image-to-video \
--param images=./social-frame-${i}.jpg \
-o social-${i}b.mp4
done
# 6 variants generated. Pick the best 3 to post.
Result: 6 video variants from 3 stills, generated in minutes. Fast models keep the iteration loop tight.
Why this pairing: Nano Banana 2 for speed (volume of stills), Seedance 2.0 Fast for speed (volume of clips). This pipeline prioritizes quantity so you can A/B test.
Pipeline 3: Design-to-Motion (Creative Exploration)
Goal: Take a design reference and explore how it would look in motion.
# Step 1: Generate a design-heavy still
anycap image generate \
--prompt "geometric abstract shapes in coral and navy, overlapping with varied opacity, editorial design style, high contrast" \
--model flux-kontext-max \
-o design-frame.jpg
# Step 2: Explore motion with Kling 3.0 (best camera dynamics)
anycap video generate \
--prompt "shapes drift apart slowly, camera orbits the composition, one shape pulses with light" \
--model kling-3.0 \
--mode image-to-video \
--param images=./design-frame.jpg \
-o design-motion-1.mp4
# Step 3: Try a different motion style
anycap video generate \
--prompt "fast zoom through the shapes, kaleidoscopic rotation, energetic pace" \
--model kling-3.0 \
--mode image-to-video \
--param images=./design-frame.jpg \
-o design-motion-2.mp4
Result: Two different motion treatments of the same still. Compare them side by side and pick the direction that works.
Why this pairing: FLUX.1 Kontext Max handles design-heavy visuals better than other image models. Kling 3.0 gives you the most expressive camera control. Together, they're the best pipeline for creative and design work.
When to Skip Image-to-Video and Go Direct
Image-to-video isn't always the right choice. Skip the still-image step when:
The scene doesn't have a static starting point. A drone flyover, a particle simulation, an abstract motion piece — these don't benefit from a locked keyframe. Use text-to-video directly.
Speed matters more than control. Quick social clips where "close enough" is good enough. Text-to-video with a Fast model gets you there in one step.
You want pure motion exploration. "Show me 5 different ways this concept could move" — text-to-video with different motion prompts gives you variety faster than generating 5 stills first.
The Full Stack: Text → Image → Video → Publish
The image-to-video pipeline is one piece of a larger workflow. Here's how it connects to the rest of the agent capability stack:
1. WEB SEARCH — research reference styles
↓
2. IMAGE GENERATION — create the keyframe
↓
3. IMAGE-TO-VIDEO — animate the keyframe
↓
4. DRIVE STORAGE — store the final clip
↓
5. PAGE PUBLISH — embed the video on a published page
Your agent can run all five steps in one session. No context switching. No separate tools. This is the full creative pipeline for coding agents — and it's only possible because all the capabilities live behind one runtime.
FAQ
Which image model gives the best starting frame for video?
Seedream 5 for quality. Nano Banana Pro for revision-heavy workflows. Nano Banana 2 for speed. FLUX.1 Kontext Max for design-heavy visuals.
Can I use the same prompt for image and video?
No — and that's the point. The image prompt describes the scene (composition, lighting, subject). The video prompt describes motion (camera movement, element animation, transitions). Keep them separate for the best results.
How do I ensure the video quality doesn't degrade from the still?
Use a quality-matched pairing. Seedream 5 → Veo 3.1 or Seedance 2.0 preserves fidelity. Nano Banana 2 → Seedance 2.0 Fast works but expect some quality tradeoff. Fast models prioritize speed over fidelity.
Can I batch-generate image-to-video?
Yes. Loop the image generation step to create multiple keyframes, then loop the video generation step to animate them. This is the social content batch pipeline described above.
Do I need to install anything separately for image-to-video?
Not with AnyCap. anycap image generate and anycap video generate --mode image-to-video use the same CLI, same auth, same runtime. No separate integrations.
The Bottom Line
Text-to-video gives you motion. Image-to-video gives you control. The two-step pipeline — generate, evaluate, animate — produces output you can actually use in production because you approved the frame before committing the motion budget.
The model pairing matters. Seedream 5 + Veo 3.1 is the premium pipeline. Nano Banana Pro + Seedance 1.5 Pro is the revision-to-motion pipeline. Nano Banana 2 + Seedance 2.0 Fast is the speed pipeline. Pick based on whether quality, consistency, or throughput matters most for your workflow.
→ Give your coding agent the full image-to-video pipeline — one CLI, all models
📖 What to Read Next
- How to Generate Video with Claude Code: The Complete 2026 Guide — The full guide to video generation for Claude Code, Cursor, and Codex.
- How to Generate Images with Claude Code (2026): 3 Methods — The image generation companion: MCP, DIY API, and AnyCap approaches.
- Best AI Video Models for Coding Agents Compared — Veo 3.1 vs Seedance 2.0 vs Kling 3.0 vs Sora 2 Pro: head-to-head.
Related Articles
- Image Generation for AI Agents: The Developer's Guide — Seedream 5, Nano Banana Pro, FLUX.1 Kontext Max, and the full image model catalog.
- What Is a Capability Runtime? — The infrastructure layer that bundles image, video, search, and storage into one CLI.
Written by the AnyCap team. We build the capability runtime that lets your agent generate images, animate them into video, and publish the result — all through one CLI.