I spent last week going through the models that are actually live on imgmov right now, not the ones we used to advertise. That sentence matters more than it sounds. Our old guide listed Agnes Video 2.5 Flash, Seedream 5.0 Lite, Seedance 2.5, and several prices that no longer describe the product. I rewrote this after checking the current generation config and pricing copy, because the fastest way to confuse someone is to hand them a beautiful guide built on a dead model list.
What Is AI Generation?
AI generation is basically using machine learning models to create visual content from text descriptions, reference images, or other inputs. The most common types include
- Text-to-Image: Generate images from text prompts
- Text-to-Video: Create videos from text descriptions
- Image-to-Video: Animate a static image into a video
- Image-to-Image: Transform an existing image using AI
What I use now
I think in three layers.
Guest trial. Agnes Image 2.5 Flash and Agnes Video V2 are the two Agnes trial models. Agnes Image 2.5 Flash is useful for checking composition and color before spending a real credit. Agnes Video V2 handles text-to-video and image-to-video. The trial is limited to three watermarked previews per guest per day, 1K quality, one result at a time, and no commercial delivery. I would use it to answer one question only, does the idea deserve more credits?
Production image. Seedream 5.0 Lite and Nano-Banana 2 both cost 5 credits per image and support eight aspect ratios. GPT Image 2 starts at 2 credits for Standard 1K and scales through a quality and resolution matrix up to 80 credits for High 4K. When a composition needs reasoning and a wider frame, I pick Seedream 5.0 Lite. When I need a photorealistic social scene quickly, I pick Nano-Banana 2. When labels or packaging text must survive, GPT Image 2 is worth the extra credits.
Production video. Seedance 2.5 is billed by resolution, 18 credits per second at 480p and 39 credits per second at 720p. Kling v3 Omni is 39 credits per second, or 49 credits per second with audio. Veo 3.1 is 80 credits for 4 seconds, 125 for 6, and 165 for 8. These are not casual costs. They are the reason I stopped telling people to generate first and think later.
The workflow I actually repeat
The quality of your AI generation depends heavily on the prompt you write. Here are the key principles.
Be Specific
Instead of 「a dog」, I write something like this.
A golden retriever running through a field of sunflowers, low tracking camera, golden hour light, shallow depth of field, natural grass movement, cinematic colorThen I cut it again. Video models punish vague motion more than image models punish vague subjects. If I cannot tell what changes between second two and second five, the model cannot either.
For images, I lock the product or subject before touching the scene. If the source photo has a ceramic mug with a handle on the left, the prompt still says handle on the left. Otherwise I get a beautiful fake product.
What I got wrong
I used to treat price as an afterthought. That was a mistake. A 5-second Seedance 2.5 clip at 720p is 195 credits before you have any guarantee it matches the idea. So now I split testing and production.
- Use the Agnes guest trial or a cheap image draft to find composition.
- Turn one approved image into a short clip.
- Only then move to Seedance 2.5, Kling v3 Omni, or Veo 3.1.
It feels slower. It is not. I stopped paying credits to discover that a prompt was wrong.
FAQ
Q: What is the best beginner workflow?
A: Start with Agnes Image 2.5 Flash for composition, Agnes Video V2 for motion, and move to premium models only after the concept survives. The trial has three watermarked previews per day, so use them deliberately.
Q: Which model should I choose for readable text?
A: GPT Image 2 is the safest starting point because it handles text and packaging better than the other image models on imgmov. Check labels carefully anyway.
Q: How much should I plan to spend?
A: Price the exact model, duration, quality, and resolution first. Seedance 2.5 and Kling v3 Omni are billed per second, while Veo 3.1 uses duration buckets. A good clip is cheaper than five random ones.