The Generative Media Landscape
Map the image, video, audio and 3D tools - and understand why diffusion models behave nothing like the chatbots you already know
Generative Media Is a Different Kind of AI
Levels 0-5 were about language models: you type, they write back. This level is about media models that produce pictures, video, voice and music. The skills transfer less than you'd expect - prompting an image model is closer to briefing a photographer than instructing an assistant.
Why Diffusion ≠ LLMs
Most image, video and music generators are diffusion models. They start from pure random noise and remove it step by step until an image (or frame, or audio spectrogram) emerges, guided by your text prompt. LLMs instead predict the next token, one at a time, left to right.
That architectural difference explains almost every quirk you'll hit:
- Prompts are descriptions, not commands. "Don't add a hat" often adds a hat - negation is weak. Use negative-prompt fields instead.
- Same prompt ≠ same image. A random seed drives each result. Lock the seed to reproduce or iterate on a specific image.
- Text and counting are hard. Diffusion struggles with legible words and exact quantities ("exactly five apples") - though 2026 models are far better.
- Iteration beats one perfect prompt. Generate four, pick the closest, refine with variations or reference images.
The 2026 Modality Map
Four families of generative media, each with its own leading tools. Names change fast - treat this as a snapshot of mid-2026, and always re-check before you commit a budget.
IMAGE GPT Image 2 (OpenAI) - best all-rounder, strong text + editing Midjourney V7 - artistic/stylized quality leader FLUX (Black Forest Labs)- photorealism leader Adobe Firefly - commercially safe (licensed training data) Ideogram 3 - best for text/typography inside images VIDEO Google Veo 3.1 - top all-rounder, native audio, 4K Kling 3.0 - cinematic motion, multi-shot storyboards Runway Gen-4.5 - granular creative control (camera, motion brush) (Note: OpenAI's Sora app is being retired in 2026 - don't build on it) AUDIO - SPEECH ElevenLabs - voice cloning + text-to-speech quality leader AUDIO - MUSIC Suno v5 - best full songs w/ vocals (licensing unsettled) ElevenLabs Music - licensed training data, clean commercial terms 3D / DESIGN Image-to-3D mesh tools, Canva AI, Adobe Firefly in-app
How You Actually Access These Tools
There are four delivery models, and the right one depends on volume and control:
Midjourney, Runway, Suno, ElevenLabs all have browser UIs. Best for learning, one-offs, and visual iteration. Usually subscription + credits.
Canva, Adobe Express/Photoshop and Google Slides embed generation directly. Best for non-technical users who already live in those tools.
Call the model from code to generate hundreds of assets. Priced per image or per second of video. This is Level 4 territory applied to media.
Run FLUX or Stable Diffusion locally via ComfyUI on your own GPU. No per-image cost, full privacy, steep setup. See the Local AI tools page.
Understanding the Cost Model
Media generation is priced very differently from chat. Knowing the unit cost prevents nasty surprises.
Images ~ per image, or a monthly credit pool
(a "fast hour" / credit ≈ a batch of 4)
Video ~ per SECOND of output - the expensive one
Veo 3.1 ~ $0.15/sec (fast mode)
Kling 3.0 ~ $0.10/sec
=> a 10-second clip can cost $1-$8 depending on model/quality
Voice ~ per 1,000 characters of text (TTS)
Music ~ per song / per generation, or a monthly quotaChoosing a Tool: Four Trade-offs
Every selection in this level comes down to weighing four things:
- Quality - does the output look the way you need?
- Control - can you nail a specific result, repeatedly?
- Commercial safety - are you legally clear to sell/publish it? (Lesson 54)
- Cost - per asset and at the volume you need.
Hands-On: Same Prompt, Two Models
Generates media by removing noise step-by-step, guided by a text prompt - the basis of most image/video/music tools
The random starting value; lock it to reproduce or iterate on a specific result
A separate field listing what to exclude - more reliable than saying "don't" in the main prompt
Web app, in-app (Canva/Adobe), API, or local/open-source
Charged per second of output - by far the costliest modality
The four trade-offs behind every tool choice; no tool wins all four
Lock your shot in cheap image tools before spending on video generation