Running a multi-step AI pipeline on Theta — text, image, vision

A follow-up to my earlier MCP test. The first one was a single image generation. This one is a chained pipeline: text generation → image generation → vision interpretation, with three different Theta models running through the MCP server in a single conversation.

No glue code. No API juggling. Just one prompt to Claude.

The prompt

Here's the exact prompt sent to Claude:

I want to test the Theta EdgeCloud MCP server's ability to orchestrate a multi-step pipeline using three different Theta models. Use the Theta MCP tools for each AI step — don't do the AI work yourself.

Pipeline:

Step 1 — Text generation
Use GPT OSS 120B (gpt_oss_120b) on Theta to write a short, vivid 2-3 sentence description of what Theta Network's decentralized GPU infrastructure looks like as a physical metaphor. Imagine it as a landscape, a city, a creature, or some other visual concept. Make it visually rich and specific.

Step 2 — Image generation
Take the description from Step 1 and craft an image generation prompt from it. Make the prompt detailed enough to produce a striking image (style, mood, lighting, perspective). Then use FLUX (flux) on Theta to generate the image.

Step 3 — Vision / image interpretation
Take the image generated in Step 2 and pass it to LLaVA (llava) on Theta. Ask LLaVA to describe what it sees in the image, in its own words.

Step 4 — Comparison and summary
Give me a clear summary that includes:
- The exact text generated in Step 1 (by GPT OSS 120B)
- The image prompt used in Step 2
- The image generated in Step 3 (display it)
- LLaVA's description of the image (Step 3 output)
- A short comparison: how does LLaVA's interpretation match or diverge from the original text in Step 1?
- Which Theta model was used for each step
- Approximate time each step took

Don't combine steps. Don't do the writing, prompting, or vision interpretation locally — route every AI generation step through the Theta MCP server. I want to see the full orchestration in action: text → image → vision → comparison.

That single message is what kicked off everything below.

The pipeline

Step 1 — GPT OSS 120B

Prompt: write a short, vivid 2-3 sentence physical metaphor for Theta's decentralized GPU infrastructure.

Output:

"A glittering megacity of glass towers rises from the horizon, each spire a GPU humming with electric veins that pulse in sync like neon arteries. Between them, luminous fiber-optic skyways weave a constantly shifting lattice, letting data torrents cascade like waterfalls of light across the skyline. At night the whole landscape glows as a single, breathing organism, each node sharing its heat and power to keep the city alive."

Time: ~2 seconds.

Step 2 — FLUX

The text from Step 1 was used to construct an image prompt with explicit style, lighting, and perspective directives. FLUX rendered at 1024×1024, 28 steps, guidance 3.5.

Time: ~9 seconds.

The result is the featured image at the top of this post.

Step 3 — Vision interpretation

This is where the test got interesting.

The plan was to send the image to LLaVA and ask it to describe what it sees. LLaVA is a conversational vision model — it can interpret an image and discuss it.

LLaVA returned state: created and stayed there. No instance available. Theta's decentralized GPU allocation works on operator availability, and at that moment no node was running LLaVA.

I switched to BLIP — a different vision model on Theta — and got an immediate response.

BLIP's description of the image:

"futuristic city in the middle of the night"

That's the full output. BLIP is an image captioner, not a conversational vision model. It returns short, literal labels.

Time: ~3 seconds.

Comparison: GPT's text vs BLIP's reading

BLIP captured the surface — futuristic city at night — but lost every layer of metaphor. That's not a Theta limitation. It's BLIP's design. For a richer reading, LLaVA would have been the right model. It just wasn't available at that moment.

Three observations

Multi-step pipelines work. Three different models, three modalities, all routed through MCP from a single conversational prompt. No code written by me.

"No instances available" is a real constraint. Decentralized GPU networks don't guarantee model availability the way AWS does. For production pipelines, fallback logic is required. I did this manually here. A real application would need to handle it programmatically.

Model choice matters more than people realize. BLIP and LLaVA both fall under "vision," but they're built for different things. Picking the right Theta model for a step is part of the design work — not something MCP solves for you.

What this would look like at scale

The full pipeline — three model calls, three modalities — registered as roughly $0.01 in EdgeCloud credit. The exact figure is below the resolution of the dashboard, but the order of magnitude is clear: fractions of a cent per run.

A media company indexing 10,000 images with this flow runs around $100 in compute. The pipeline composes. Once it works for one, it works for ten thousand — same MCP orchestration, no rewriting.