Building a diffusion model that spans image editing and temporal keyframes introduces complexity across every layer of the stack. The model must handle large parameter counts, long diffusion sequences and high-resolution data, while running reliably across multi-node GPU clusters for extended training cycles.

At this scale, long-running jobs are vulnerable to hardware interruptions, node failures, networking instability and I/O bottlenecks, any of which can terminate a run and erase days of progress. These are the realities of large-scale training.

At the same time, Higgsfield AI’s broader mission gained momentum. The company is moving towards making creativity accessible for both professional social media creators and the largest advertising agencies and global brands. That means the training pipeline must support rapid iteration, consistent quality and predictable delivery.

To meet these demands, Higgsfield AI collaborated with Nebius as a hands-on infrastructure collaborator deeply involved in solving the challenges of large-scale training. From day one, the joint work focused on removing the systemic risks that derail long training runs like unreliable capacity.

As Alex Mashrabov, Founder and CEO of Higgsfield AI, explains: “From day one, Nebius felt like a co-engineer, not just a cloud provider. Their infrastructure, paired with the new Blackwell GPUs, gave us the confidence to push the boundaries of generative AI — enabling faster iterations, stable scaling and exceptional reliability across model development, training and inference.”

This co-engineering mindset proved critical as Higgsfield AI scaled model complexity and business demand.

The computational demands of Higgsfield AI’s training pipeline required state-of-the-art hardware. The team deployed NVIDIA HGX B200 systems, ramping up to 180GB of GPU memory per device to support large model parameters and high-resolution batches without resorting to aggressive memory workarounds that slow training.

Higher memory bandwidth and strong FLOPS performance further reduced pressure points common in diffusion workloads, allowing the team to design the training stack around predictable hardware behavior.

The team designed the training stack around the hardware, using Blackwell’s strengths to simplify other parts of the system.

Memory optimization beat recomputationMemory optimization beat recomputation

To manage GPU memory pressure, Higgsfield AI evaluated two primary strategies: activation checkpointing and distributed optimizers. Activation checkpointing reduced memory usage but introduced significant recomputation overhead during the backward pass, slowing training to an unacceptable degree.

The team observed that while checkpointing saved memory, it increased per-iteration time enough to negate its benefits. Distributed optimizers, by contrast, effectively managed memory without imposing the same performance penalty. Higgsfield AI opted to rely solely on distributed optimizers for the full training run.

The takeaway: in large diffusion models, conserving memory is important — but not at the expense of iteration speed.

FlashAttention-based optimizations improved throughputFlashAttention-based optimizations improved throughput

The team benchmarked multiple implementations, including FlashAttention variants (designed for Hopper) and cuDNN fused attention (built for Blackwell). Empirical testing revealed that a hybrid configuration maximized throughput for Higgsfield AI’s diffusion workload. This tuning step was essential for sustaining performance across long training runs.

Kernel fusion compounded at scaleKernel fusion compounded at scale

To extract additional efficiency, Higgsfield AI leveraged torch.compile() to JIT-compile graph-friendly sections of the model. This enabled kernel fusion, reducing memory access overhead by combining multiple operations into single GPU instructions.

While the gains per iteration were modest in isolation, they compounded across longer runs, saving thousands of GPU hours over the full training lifecycle.

Training stability wasn’t just a problem of accessing enough compute at the right time. Higgsfield AI designed a multi-stage data curriculum that progressed along two dimensions: resolution and data quality.

Early training focused on mid-resolution imagery (around 720p) to teach the model basic structure. As training advanced, resolution increased beyond 2K, allowing the model to learn fine-grained detail and texture. In parallel, the dataset evolved from massive, broadly sourced data to increasingly curated, high-quality samples. For an audience of fast-moving creatives, this reliability and aesthetic focus is the difference between frustration and execution.

Nebius played a critical role by enabling high-throughput data distribution and storage architectures tuned for streaming large datasets. Higgsfield AI decoupled data preprocessing from training: raw image-caption pairs were filtered, encoded into latents and cached by resolution in a shared, location-aware store. Training nodes then pulled pre-processed batches asynchronously, ensuring GPUs were never idle waiting for data.

Once the training pipeline stabilized, Higgsfield AI turned to alignment. High-quality outputs need to be more than technically correct. They had to reflect human aesthetic judgment with expression and visual appeal.

The team used Direct Preference Optimization (DPO), generating multiple candidate outputs per prompt and ranking them with human evalautors. This process required reliable access to large GPU clusters to support repeated generation and retraining cycles.

Nebius’ stable multi-node clusters made it possible to run these preference-driven training loops continuously.

By solving the hardest training problems first, Higgsfield AI built a foundation that scaled with the business. Training runs completed reliably. Iteration cycles shortened. Engineering effort shifted from firefighting to refinement.

That stability translated directly into speed. Higgsfield AI moved from zero to $100M in seven months, then crossed a $200M run rate faster than any other gen AI company — in just two months, soon after without re-architecting its training stack or slowing development.

As Mashrabov puts it: “Training a multi-billion-parameter diffusion model isn’t just about compute — it’s about stability, bandwidth and trust. Nebius delivered all three.”

For Higgsfield AI, the collaboration with Nebius delivered more than infrastructure. It was a nimble co-engineering collaborator that enabled the company to push the boundaries of generative AI while keeping creators at the center of the experience.