FLUX 2 And HunyuanVideo 1.5: When Open AI Starts To Feel Uncomfortably Real

There is a point where visual AI stops feeling like a toy and starts to quietly unsettle you. FLUX 2 from Black Forest Labs and Tencent’s HunyuanVideo 1.5 sit right on that line, where open models begin to match the polish and control that used to be reserved for big closed systems.

In this breakdown, we will look at what FLUX 2 does for image generation, how HunyuanVideo 1.5 changes open video, and why both matter if you care about real production work, not just fun demos. The goal is simple: help you understand what these models can actually do today, where they still miss, and how you can start using them without melting your GPU.

---

FLUX 2: Open-Weight Images That Feel Uncomfortably Real

When Black Forest Labs dropped FLUX 2, the first reaction across timelines was almost disbelief. Faces held together across shots, hands looked normal, and characters stayed consistent in scene after scene. It felt like someone had taken the best parts of closed models and quietly handed them to the open world.

At a high level, FLUX 2 is designed to close the gap between open weights and the polished systems used in production. You are not just getting “good for open source” quality. You are getting images that can sit next to commercial tools without obvious tells.

The core pitch comes down to three big wins:

• Multi-reference consistency. You can feed up to 10 reference images into FLUX 2 and ask it to keep characters, products, or a style consistent across new generations. That means the same model can keep your lead character on brand through key art, social crops, and variant poses.
• Production-ready detail. FLUX 2 works at up to 4 megapixels, so you get real resolution to work with, not just upscaled noise. Skin looks stable, lighting behaves in a believable way, and hands largely stop being a punchline.
• Control without LoRA chaos. Instead of juggling separate LoRAs or duplicating checkpoints, multi-reference behavior and text-based editing live inside the core model. You spend more time directing, less time debugging.

For many art directors, the first real shock is how stable style and lighting feel. You can set up a “visual language” for a project and expect FLUX 2 to stay inside those constraints across a series, which used to be a clear advantage for closed tools.

If you want to see how Black Forest Labs positions it themselves, their own overview of FLUX.2 as frontier visual intelligence is worth a careful read.

Text Rendering That Finally Looks Client-Ready

Text has been the last big weak point for image models. Slide decks, UI mockups, and product shots with packaging always looked a bit warped, even when everything else was sharp.

With FLUX 2, that changes in a noticeable way. The model treats typography as a first-class part of the image, not an afterthought.

That matters if you build:

• Infographics and diagrams that need legible labels.
• UI and product flows where layout hierarchy must stay intact.
• Logos, posters, and thumbnails with brand text that cannot be half-melted.
• Social graphics that go straight from prompt to client review.

You will still clean things up in Figma or Photoshop if the stakes are high, but for a lot of work FLUX 2 gets you surprisingly close on the first pass. For many teams, that alone saves hours per week.

---

Inside FLUX 2: Why The Architecture Feels Different

Black Forest Labs did not just increment the old stack. They rebuilt FLUX 2 around a hybrid design that splits semantic understanding from visual synthesis, which is a big reason it behaves more like a careful illustrator than a random noise machine.

At a simple level, you can think of FLUX 2 as two main parts that talk to each other.

1. Mistral 324B vision-language model

   This part reads your prompt and your reference images together. It learns what belongs where, how objects relate in space, how materials should react to light, and what structural elements must stay consistent.

   In practice, this is why FLUX 2 can understand a prompt like “same character from reference, now in a rainy cyberpunk street, lit only by neon reflections” and still keep identity and pose under control.

2. Rectified flow transformer

   This is the image engine. It is responsible for composition, shapes, edges, textures, and all the small details that make the final frame feel grounded. The rectified flow setup helps the model move more directly from noise to a coherent image, which reduces some of the drift you see in older diffusion pipelines.

The two halves are glued together by a custom VAE that Black Forest Labs trained from scratch. Instead of reusing a generic encoder, they tuned it to balance three things at once:

• How easily the model can learn.
• How compact the latent representation is.
• How well the decoded image matches the internal representation.

The technical phrasing can sound abstract, but the effect is concrete: you see sharper detail, fewer strange compression artifacts, and a more stable latent space for editing and inpainting.¹

For those who want to inspect the weights, the FLUX.2-dev model card on Hugging Face lays out the specs, constraints, and intended uses in useful detail.

---

FLUX 2 Variants: Choosing The Right Power Level

FLUX 2 is not a single monolithic release. Black Forest Labs ships a small family of variants tuned for different use cases, from heavy production work in the cloud to experimentation on local GPUs.

Here is the lineup in plain terms.

• FLUX 2 Pro

  This is the flagship version that lives behind Black Forest Labs’ own playground and API. It is tuned to compete directly with closed production systems, with a focus on reliability, speed, and polished defaults.

• FLUX 2 Flex

  This version opens up more knobs. If you like adjusting sampling steps, guidance scales, or trading speed for a bit more detail, Flex is where you experiment. It is the “power user” build.

• FLUX 2D

  This is the 32 billion parameter open-weight checkpoint that has the community excited. It handles both text-to-image and image editing in a single model, which keeps pipelines simple.

  You can run FLUX 2D through providers like FAL, Replicate, Runware, Vera, Together AI, Cloudflare, or Deep Infra, or run it locally if your GPU has the VRAM. Nvidia has also highlighted FLUX 2’s FP8-optimized pipeline for ComfyUI and RTX cards, which lowers memory needs and speeds up inference.

• FLUX 2 Klein (coming soon)

  Klein is a distilled variant that will ship under Apache 2. The idea is a smaller model that still captures the behavior people care about from the larger checkpoints, but is easier to host and embed into products.

The nicest part is that all variants share the same core skill set: multi-reference control, strong text handling, and text-based editing. You do not have to juggle separate checkpoints for generation and editing, which simplifies both pipelines and mental models.

For a clear feature overview, the dedicated FLUX.2 product page gives an up-to-date matrix of capabilities and intended scenarios.

---

FLUX 2 Benchmarks: Stress Tests, Not Just Pretty Samples

Benchmarks are often skewed, but in FLUX 2’s case they tell a consistent story. Across public ELO-style evaluations, FLUX 2 scores near the top for prompt following, global composition, and overall visual quality, while keeping inference cost at a level that is friendly for real workflows.

One of the best illustrations is a deliberately chaotic prompt used to stress spatial reasoning:

A monkey holding a pink banana while sitting on a tiger, with a horse riding an astronaut in the background.

Many models will grab the tokens, then mix them into visual soup. Limbs fuse, objects overlap in impossible ways, and the “in the background” constraint is ignored. FLUX 2 manages to keep the scene structured: the monkey is on the tiger, the banana is pink and visible, and the horse and astronaut stay in the distance where they belong.

This is not about a single meme image. It is about whether a model can respect relationships in complex shots, which is critical for storyboards, product scenarios, and any multi-object setup.

If you want a more formal view, the Diffusers integration overview for FLUX 2 walks through metrics, default settings, and how the model behaves inside the standard open-source tooling.

Inline image prompt: surreal but coherent 4K scene of a monkey holding a pink banana while sitting on a tiger, with a horse carrying an astronaut in the far background, bright natural lighting, playful but realistic style.

The larger context here is Black Forest Labs’ open-core stance. They ship production-grade capabilities in their own playground and API, while releasing strong open weights that fit cleanly into tools like Diffusers and ComfyUI. That combination is what builds trust in the developer community.

---

HunyuanVideo 1.5: Open Video That Finally Moves Like Video

While FLUX 2 was shaking up image generation, Tencent quietly raised the bar for open video with HunyuanVideo 1.5. On paper, it sounds almost modest: an 8.3 billion parameter model that fits on consumer GPUs. In practice, the results feel far from modest.

The two main pain points for open video have been clear for a while:

• You needed huge VRAM to get anything decent.
• Motion, physics, and frame-to-frame stability would fall apart under stress.

HunyuanVideo 1.5 tackles both at once. It delivers controlled motion, smooth camera paths, and surprisingly robust temporal consistency, all from a model size you can run at home. On top of that, it ships uncensored, which matters for many research and creator workflows.

Tencent’s own HunyuanVideo 1.5 GitHub repository lays out the architecture and releases both code and weights for experimentation.

What HunyuanVideo 1.5 Actually Does Well

HunyuanVideo 1.5 comes in two primary base variants, one for 480p and one for 720p output, both tied to a dedicated super-resolution system that upscales cleanly to 1080p. The upsampler works directly in latent space, which avoids the shimmering interpolation artifacts that have plagued older pipelines.

The strongest traits show up across three areas.

• Motion and stability

  Movement feels grounded. A figure skater spins with weight and balance. A character walks through a space without limbs dissolving. Objects keep their shape over time instead of melting.

• Instruction following

  HunyuanVideo 1.5 reads long, detailed prompts and translates them into camera moves, lighting changes, and multi-step actions. You can describe “slow dolly in, then tilt up as the couple kisses and text appears above them” and the model will attempt that full sequence, not just a static frame.

  It does this in both English and Chinese, thanks to a text encoder built for rich multimodal alignment.

• Image-to-video consistency

  When you seed it with a still frame, HunyuanVideo 1.5 keeps the look remarkably close over the clip. The “cinematic woman turning her head from a retro portrait” example is a good reference: color, texture, and vibe stay intact as motion is added.

For a quick feel of how this translates to real footage, Tencent’s HunyuanVideo online demo page includes curated prompts and clips that show the model under different kinds of stress.

Concrete Examples From The Demos

It helps to anchor capabilities in actual scenes. Some of the clearest examples from the official tests include:

1. Figure skater

   A skater spins on the ice with fluid, believable body mechanics. The motion holds up across frames, with no major tearing or limb glitches.

2. DJ close-up

   A DJ moves slightly as the camera lingers on their face. Micro-expressions and head movement look natural, not robotic.

3. Cinematic B-roll

   Slow, steady shots of bread on a marble slab or leaves covered in dew highlight how the model handles subtle texture, lighting, and shallow depth of field.

4. Physics-heavy action

   In a soda can crushing test, the can deforms in a way that feels roughly physical instead of snapping between unrelated shapes.

5. Camera motion

   One standout is a shot where the camera pans down to follow a cat and then changes focus, or another where the camera pulls back and rises to reveal a wide desert scene. The instructions for pan, tilt, zoom, and orbit are respected with a level of discipline that has been rare in open models.

6. Stylized worlds

   The system moves beyond realism into anime, retro, and claymation. The clay “cake man eating himself” clip is strange, but stylistically consistent and coherent from frame to frame.

7. Image-to-video portrait

   Seeded with a still portrait, a woman turns her head slowly while lighting and tone stay anchored to the original design.

These are not flawless outputs, but they are coherent in a way that makes editing and grading feel worth the time.

---

HunyuanVideo 1.5 vs Open-Sora 1.22: Where It Pulls Ahead

Open-Sora 1.22 has been a strong reference point for open video. Comparing it directly with HunyuanVideo 1.5 reveals where Tencent’s work pulls ahead and where both still struggle.

Some of the clearer differences from side-by-side tests:

• Chaotic market with explosions

  Asked for an orbiting camera around a man in a chaotic market, HunyuanVideo 1.5 actually performs a visible orbit. Open-Sora tends to keep the camera more static.

• Figure skating and parkour

  In anatomy-heavy motion like skating or parkour, both models run into edge cases, but HunyuanVideo keeps body structure more coherent under fast movement.

• Dimension-bridging ladder

  In a surreal scene where a ladder links two visual domains, HunyuanVideo blends both sides in one continuous space, while Open-Sora treats them more like a hard cut between scenes.

• Romantic push-in with text

  On a couple kissing as the camera pushes in and tilts up for overlaid text, HunyuanVideo gets the camera move closer to the prompt but still struggles with clean text. Open-Sora misses both camera instruction and legible text.

• Action and “jiggle physics”

  In fast action sequences, HunyuanVideo usually delivers more energetic motion but slightly messier hands. Open-Sora moves slower but keeps structure a bit cleaner. Both handle influencer-style talking head videos reasonably well, although pacing often needs tightening in post.

• Character identity

  Neither model reliably recognizes specific IP characters like Naruto or One Punch Man, which is expected without custom LoRAs.

In human evaluations, HunyuanVideo 1.5 tends to win for instruction following, visual richness, and motion effects, while Open-Sora earns points for structural stability and frame-to-frame consistency. On balance, users showed a clear preference for HunyuanVideo outputs in both text-to-video and image-to-video setups.

The weak area for both is still text. Clean titles, lower thirds, and in-scene signage remain a challenge, similar to where image models were not long ago.

---

Under The Hood: Why HunyuanVideo 1.5 Runs On Real GPUs

The architecture behind HunyuanVideo 1.5 is more sophisticated than the parameter count suggests. It is tuned to respect three constraints at once: quality, temporal stability, and hardware cost.

The backbone is a unified diffusion transformer that works together with a 3D causal VAE codec. That VAE compresses spatial information at a 16:1 ratio and temporal information at 4:1, which keeps video latents compact while still letting the model reconstruct sharp, detailed frames.²

To stop long clips from overwhelming GPUs, Tencent added a selective and sliding tile attention system (SSTA). In plain terms, SSTA pays full attention only where it matters, then slides that attention window over space and time instead of trying to optimize the whole clip at once. You keep coherence without paying for unnecessary context.

Instruction understanding is handled by a multimodal large model paired with a dedicated text encoder. This is what lets prompts drive camera language, action sequences, and lighting changes with more control than a simple text encoder would allow.

Training used a multi-stage setup, from pretraining to post-training, with an optimizer stack tuned for faster convergence and better motion coherence. The result is a model that “feels” more stable than its size would suggest.

If you want to reproduce the local setup, the HunyuanVideo 1.5 ComfyUI tutorial provides a clear step-by-step workflow.

How To Run HunyuanVideo 1.5 Locally Without Cooking Your GPU

Running HunyuanVideo 1.5 at home is less painful than earlier generations, especially if you use ComfyUI.

A simple path looks like this:

1. Update ComfyUI to the latest version.
2. Download the text encoders, diffusion model, and VAE from the official sources.
3. Drop the files into the correct model folders, then refresh your ComfyUI model list.
4. Load the provided HunyuanVideo workflow, set your prompt and seed image if needed, then start generation.

If your VRAM is limited, shift to FP8 or GGUF variants. The smallest GGUF build is under 5 GB and still runs full image-to-video generation, with some tradeoffs in fine detail. You usually only need to swap the UNet loader node and leave the rest of the graph unchanged.

What This All Means For Creators And Teams

Taken together, FLUX 2 and HunyuanVideo 1.5 mark a clear moment for open visual AI. On the image side, FLUX 2 gives you production-grade quality with the kind of control and consistency that used to lock you into closed ecosystems. On the video side, HunyuanVideo 1.5 shows that smooth, instruction-aware motion is no longer reserved for massive proprietary stacks.

For solo creators, this means you can storyboard, design key art, and generate test footage on consumer hardware, then refine in your regular tools. For teams, it means you can build pipelines around open components, not just third-party black boxes, and still hit a quality bar that clients will accept.

If you are already comfortable with diffusion workflows, the path into FLUX 2 is straightforward, especially through Diffusers and ComfyUI. For video, HunyuanVideo 1.5 gives you a practical way to add motion to your stack without a new line item in your cloud bill.

The pace here is fast, and it will keep increasing. The steady trend is clear: open models are no longer just “good enough.” In more and more workflows, they are becoming the default.

The quiet shift is this: reality in your projects is now something you can dial in, not just capture.