Seedream 5.5 Transparent PNG: Native RGBA, Glass Gap

Of Seedream 5.5's three upgrades, native RGBA is the most practical: direct output of transparent PNGs with an alpha channel, so assets go into design workflows without cutouts. It's also one of the clearest differentiators from GPT Image 2.6. But three independent tests found the same limitation: glass, water and other see-through objects themselves come out essentially opaque. This article explains what native RGBA can and can't do.

What native RGBA means

5.5 outputs images with an alpha channel directly — fully and semi-transparent — with edges and transparent areas handled by the model, no post-processing cutouts or edge cleanup needed. Transparent assets can be generated from a text description or a reference image. Background modes: auto, transparent, opaque. Transparent output is always PNG — JPEG cannot carry an alpha channel.

The difference from the previous generation is fundamental: 5.0 Pro's transparency option only worked when editing a single image that already had an alpha channel — it could preserve transparency, not create it. 5.5 can generate transparent PNGs from text alone. That's why the word "native" matters: transparency isn't post-processing, it's what the model outputs directly.

Hands-on: clean cutouts

Quantslant's two rounds are the most systematic transparency tests: round one, transparent backgrounds 4/4 with a genuine alpha channel; round two, 10/10 genuine RGBA PNGs. Hair and mesh bags cut cleanly; smoke 2/2 genuinely semi-transparent. Seedance.io's sticker cutout test was equally clean. In ifanr's test, layer decomposition returned structured recognition results for text layers, and transparent PNG layers carried stacking order, position, name and description — transparency here isn't just "background removed," it's structured output.

One operational detail: ifanr's first transparency test still returned a normal background — an explicit alpha-channel request was needed to get a transparent image. Write the transparency requirement into the prompt; don't rely on the default.

Hands-on: the glass problem

The consistent limitation across three tests: cutouts are clean, but glass, water and other see-through objects themselves are essentially opaque. Quantslant: glass 0/3 see-through (round one), glass and water 0/4 see-through (round two). Seedance.io: the glass perfume bottle's background came out clean, but the bottle itself was opaque — only about 0.5% of pixels semi-transparent, all at the edges. ifanr: the first transparency test still returned a normal background. This is currently the most reproducible RGBA limitation.

It has practical consequences: if you want "a product shot on a transparent background," 5.5 handles it fully; if you want "a glass whose body is itself see-through," 5.5 currently can't — it gives you a glass with a transparent background but an opaque body. Designers should know this boundary before using it.

Layer decomposition: transparent PNG, advanced

Layer decomposition is a separate endpoint: feed in a finished image, get back a base image + transparent PNG layers (stacking order, position, name and description), with prompts to specify elements to split (e.g. "bottle, cap, label text, shadow"); text layers return structured recognition results. The layer cap is reported inconsistently: Volcano's launch notes say up to 100 layers, WaveSpeed/Atlas/ComfyUI say 99; apimodels' max_images is 2–100 (including the base). The reconciling explanation — "99 layers + base = 100 images" — is plausible but unconfirmed by ByteDance; this page says "~100 layers." For reference: 5.0 Pro gave the base + up to 16 layers.

Pika Image Studio added 5.5 on Nov 11, leading with 100+ layer decomposition for fine editing. ComfyUI Partner Nodes support the decomposition endpoint too. Layer decomposition is billed separately, from $0.60 on WaveSpeed.

Workflow: from transparent assets to design

Higgsfield's web app is the most direct entry: pick Transparent in the Background menu, 1–4 images per run, 2K max. On the API side, WaveSpeed, OpenRouter and other third parties are all live, with the background-mode parameter auto/transparent/opaque. ComfyUI Partner Nodes support transparent output. ByteDance's claim is that assets go straight into downstream design workflows, saving cutouts and edge cleanup — hands-on tests support the first half (cutouts really are clean), but the glass problem shows "straight into" has a precondition: the asset itself can't be a see-through object.

Two final notes: first, transparent output is always PNG — converting to JPEG yourself loses the transparency. Second, commercial rights: the launch materials say nothing about licensing, and API output rights follow each platform's terms; this page makes no commitment.