Seedream 5.5 NSFW: is Seedream 5.5 uncensored?

How far Seedream 5.5 goes, level by level

We ran the same five-level test we use for every image model on this site, from Mild up to Maximum, in each of the three things Seedream 5.5 can do. The table counts runs that came back as requested.

So the top level is not where it stops. Seedream 5.5 cleared Maximum in every mode, and the two misses sit one and two steps below it. That shape is common in this data: several image models struggle in the middle of the scale while clearing the top, because the levels test different kinds of request rather than one steady climb.

For scale, its Freedom Score of 97.8 ranks it among the more permissive image models we have tested, and its Spicy Index of 84.5 is the highest on our image board. That index has no public arena results behind it yet, so it may move.

Text to image, editing and layers do not behave alike

• Text to image: 18 of 18. Every level came back as requested, in 15 to 38 seconds per image. This is the most dependable route on the model.
• Editing: 16 of 18. Both misses are here. An edit also starts from an image you supply, so what you get depends on that source as well as on the request. Edits took 14 to 37 seconds.
• Layer split: 18 of 18. Layers came back complete, with nothing masked out or swapped for something milder, in 14 to 47 seconds. If your work ends in layered files, adult material does not appear to get in the way.

The main site reports 56 runs because it includes two extra layer splits that sit outside the index. Both of those worked too.

What softened means, and what you pay

A softened run is not an error. The job finishes, you get an image, and it is simply milder than the brief. That matters for budgeting, because a softened image is a completed job and is charged like any other, while a refused or failed run costs nothing. In our test no run was refused at all, so the realistic extra cost on Seedream 5.5 is an occasional edit you decide not to keep.

Four of the runs went through the live API, and each was charged exactly what it was quoted.

Seedream 5.5 against the other Seedream versions

No version in the family refused a run, so the choice comes down to how often each one softens. Seedream 5.5 has the largest sample of the six and the lowest starting price. 5.0 Lite is the only one that softened nothing, though it cannot split layers. 5.0 Pro sits level with 5.5 on share of runs delivered.

Seedream 4.5 and 5.0 Flash are the two that clearly give ground. Flash softened about one run in five, and 4.5 delivered neither of its Very strong runs, although ten runs is too few to lean on. If you are on either of those for adult work, 5.5 costs less than 4.5 up to 2K and softened far less often than Flash in our runs. For the non-adult side of that decision, see our same-prompt test of 5.5 against 5.0 Pro.

When another model fits better

• Most of your work is editing and a softened result is a real problem. Qwen Image 2.1 Pro delivered all 42 of its runs and edits as well as generates. Qwen Image Edit Spicy, an edit-only model, delivered all 23.
• You want the Seedream look with nothing softened. Seedream 5.0 Lite delivered all 34 runs, at a higher starting price and without layers.
• You need your own character or style file. Seedream 5.5 does not load LoRA files. Qwen Image 2.1 LoRA delivered all 42 runs and MiniMax H3 Image LoRA all 38.
• You only generate from text and want a low bill. Z-Image Spicy Pro delivered all 19 of its runs at a similar price, though it cannot edit.

The wider field is in the image ranking and the editor ranking.

What this test does and does not cover

We ran these tests on 7 November 2026 through the hosted API on SpicyAPI. Each level has three runs per mode, and six at Maximum, so this is a clear pattern from a small sample rather than a promise about your next request. Results can change if the model or its service changes.

Every test used adult subjects only. We do not test real or identifiable people, content without consent, or anything involving minors, and all three are outside the rules for this model. We publish level names and results, not test prompts or test images. The picture below is an ordinary public sample.