ByteDance, TikTok's parent company, has just released two models that form a formidable duo: Seedream 5.0 Pro (images, launched 8 July) and Seedance 2.5 (video). The former no longer just generates a pretty image; it breaks it down into editable layers like a design tool. The latter generates up to 30 seconds of video in a single pass, with no cuts. The message is clear: ByteDance is no longer aiming for the viral gadget, but for professional production.
A year ago, generative visual AI produced pretty images shared for fun, and short videos of a few seconds that impressed without being truly usable. In July 2026, ByteDance has just crossed a threshold with two models designed not to wow, but to replace part of a creative studio's workload. Let's break down this tandem, its real capabilities, and the big legal shadow that comes with it.
Seedream 5.0 Pro: when AI understands design, not just images
Launched on 8 July 2026, Seedream 5.0 Pro is a multimodal image generation model (multimodal means it accepts several types of input: text, reference images, sketches). But its real breakthrough isn't the raw quality of its images; it's what you can do after generating them.
The most striking feature is called layer separation. Concretely, the model can decompose an entire poster into more than ten independent, editable layers: the text, the main subject, the background, the decorations. Better still, areas hidden behind the main subject are automatically reconstructed, allowing you to move or replace an element without leaving a hole. You can swap a parrot for a peacock in a poster, and the background cleanly reappears where the parrot was hiding it.
Until now, fixing a detail in an AI image often meant regenerating everything, losing the image you liked. Seedream 5.0 Pro introduces precision editing: you select an area (point, lasso, brush) and the model only modifies that element, preserving the lighting, shadows and composition of the rest. For commercial use, it's the difference between a one-click adjustment and starting over completely. It's exactly the kind of control that consumer image generators have been missing.
The model brings three other assets designed for professional work. Complex information visualisation: it turns dense data and text into structured infographics, with charts, timelines and a clear visual hierarchy, a task where most models fail. Native multilingual text rendering in 14 languages (French, Arabic, Japanese, Korean, Russian, Thai), respecting each one's typographic rules, a notoriously difficult problem for AI. And photographic quality pushed on skin textures, physical lighting and portraits. Everything comes out in native 2K resolution, with upscaling possible to 4K, making it usable for professional print, not just social media thumbnails.
Seedance 2.5: 30 seconds of video, in a single pass
The video companion to this duo is Seedance 2.5, presented by ByteDance at its Volcano Engine FORCE conference on 23 June, with a gradual rollout in early July. Its main technical feat can be summed up in one sentence: generating a continuous 30-second video in a single pass, without stitching together shorter segments, without visible cuts or seams.
That sounds trivial, but no commercial AI video tool had natively reached this duration. Most cap out at a few seconds, ten at most, then stitch pieces together, causing consistency jumps (a character whose face changes slightly between segments). ByteDance claims to have solved this at the architecture level, via a system that keeps character identity, lighting and camera position consistent throughout the shot.
The other major new feature is input capacity: up to 50 multimodal references in a single generation (images, video clips, audio, 3D models), up from 12 in the previous version. For a brand team that already has a visual identity, campaign footage and character models, this means being able to inject everything at once so the video stays true to the visual identity without manual correction. The model natively outputs 4K with 10-bit colour depth, and co-generates synchronised audio in the same space as the image.
The real project: a complete production ecosystem
Taken separately, these two models are impressive. Together, they reveal ByteDance's real strategy. Seedream 5.0 Pro generates high-quality still images that serve as starting frames for Seedance 2.5, which then animates them into 30-second clips with consistent characters and scenes. You generate a perfect image, break it into layers to adjust it, then animate it into video. The entire pipeline, from idea to final clip, in a single ecosystem.
Here's how this duo compares, to frame the use cases.
| Model | Type | Key strength | Who it's for |
|---|---|---|---|
| Seedream 5.0 Pro | Image | Editable layers and pixel-level editing | Designers, marketers, e-commerce |
| Seedance 2.5 | Video | 30 continuous seconds, 50 references | Studios, brands, directors |
It's no coincidence that ByteDance is investing so heavily. The company already owns TikTok and CapCut (the video editing tool with 400 million monthly users), giving it colossal global distribution for these tools. Its Seedance platform for businesses is reported to have already reached $2 billion in annual recurring revenue. The logic is the same as seen elsewhere in AI: control the entire chain, from generation to sharing.
The fly in the ointment: Hollywood hasn't forgotten
It's impossible to talk about these models without mentioning the legal cloud hanging over them. As we recounted in our article on Seedance 2.0, the predecessor sparked a storm in early 2026: ultra-realistic clips featuring real actors and protected characters went viral, prompting cease-and-desist letters from Disney, Warner Bros, Paramount and Netflix, as well as a lawsuit from the Motion Picture Association. ByteDance had to suspend the global rollout and add watermarks, face-blocking filters and detection of protected characters.
These disputes remain unresolved. And every new capability that makes these models more powerful also raises the stakes of this unsettled conflict. A model capable of producing 30 seconds of coherent 4K video is a formidable tool for a legitimate creator, but also a fearsome weapon for counterfeiting if the guardrails give way. It's the permanent tension of these technologies, which we were already exploring regarding deepfakes and our relationship with images.
It needs to be said clearly for a European company: ByteDance is a Chinese company, and Chinese national intelligence law applies in principle to any data passing through its services, regardless of where the content is generated. Entrusting confidential brand visuals, unreleased product mockups or strategic campaigns to these tools is not a trivial decision. The quality is real, as is the value for money, but the question of data confidentiality deserves to be asked before any serious professional use.
What to remember
Seedream 5.0 Pro and Seedance 2.5 mark a clear shift: visual AI is moving from being a generator of fun images to a production tool capable of competing with professional creative pipelines. Layer editing, 30 seconds of continuous video, native multilingual support — everything points to the same goal: bringing generative AI into real studio and brand workflows.
That said, most of these capabilities are for now announced by ByteDance itself, without large-scale independent verification, and the dispute with Hollywood plus the question of data sovereignty weigh heavily. The technology is here, spectacular. The question is no longer whether AI can produce professional-quality content, but under what conditions, for whom, and at what cost to those whose job was, precisely, to create these images.