TWIML AI Podcast

Fatih Porikli on making image models precise, not just plausible

Fatih Porikli· VP of Technology at Qualcomm
·~56 min·English·TWIML
MultimodalInferenceAI Infrastructure
TL;DR

Qualcomm's Fatih Porikli argues that image generation's next frontier isn't a bigger model but smarter decomposition: reward the right objective, split planning from rendering, route to specialists, and tile the work in latent space so precise, high-resolution images run on a phone.

01The Core Problem

Plausible Isn't Precise

<strong>Image models learned to look real before they learned to be correct.</strong> The next frontier is closing the gap between a plausible picture and the precise one you actually asked for.

have become incredibly good at producing very realistic images. The lighting looks you know natural right the details look right and the overall quality can be amazing

— Fatih Porikli, TWIML AI Podcast
Key Insight
Benchmark-topping realism can mask total failure on intent. A model can score high on image quality while generating the wrong number of people, the wrong faces, or the wrong composition — because the metrics that made it photoreal never measured whether it obeyed you.
02The Framework

Three Buckets of Unfinished Work

<strong>Everything left to solve sorts into three buckets: controllability, quality, and efficiency.</strong> The whole conversation hangs on these three axes.

there are still many things to be done and that's why you know we publish present many papers addressing such challenges

— Fatih Porikli, TWIML AI Podcast
Key Insight
None of the three is framed as an architecture problem. Porikli's bet across every paper is that the base model is already capable enough — what's missing is the right objective, the right decomposition, and the right memory budget. That reframes better images from a scaling problem into an engineering one.
03Controllability - DisCo

The Cookie-Cutter Face Problem

<strong>Ask for a group and models paint the same face over and over — because training rewarded realism, never diversity.</strong> DisCo fixes it by making diversity itself an optimization objective.

existing training objectives focus heavily on the realism and matching the user prompt, but they don't explicitly encourage diversity

— Fatih Porikli, TWIML AI Podcast
Key Insight
The deeper lesson isn't about faces — a model silently inherits whatever its loss function measured. If nobody scored diversity, the model never learned it, no matter how much data it saw. Many capability gaps are really unmeasured objectives in disguise.
04The Turning Point

Maybe One Model Is Doing Too Much

<strong>A single model must parse the prompt, count subjects, place them, reason about their interactions, hold each identity, and render — all at once.</strong> Maybe that is too much to ask of one pass.

Maybe we are also asking a single model to solve too many difficult problems at once.

— Fatih Porikli, TWIML AI Podcast
Key Insight
This is the hinge of the entire interview. Every paper after it is a different answer to the same question — split the objective (DisCo), split the pipeline (R2Can), split across specialists (routing), or split the image into patches (PixelRush). Do less per model is the through-line hiding in plain sight.
05Decomposition - R2Can

An Architect and an Artist

<strong>Split generation into planning and rendering, the way a human artist does.</strong> An architect lays out where each subject goes; an artist renders the photorealistic image on that plan.

Instead of asking model to do everything, we separate uh planning from rendering similar to how a human artist would work.

— Fatih Porikli, TWIML AI Podcast
Key Insight
The clever part is what the architect passes down: not an image, just a few (x,y) points. The artist is fine-tuned to honor those anchors. Decomposition works here because the interface between stages is tiny — a handful of coordinates — so errors cannot compound the way they would if one blurry image were handed to the next.
06Orchestration

One Size Doesn't Fit Everyone

<strong>Instead of one model tuned for every attribute, route each prompt to the specialist that handles it best.</strong> Image generation starts to look like an agentic pipeline.

maybe size doesn't fit everyone if you really really want to generate something amazing top of the line. So we need such specialization

— Fatih Porikli, TWIML AI Podcast
Key Insight
Routing reframes the whole problem: you stop building one universal model and start building a portfolio plus a dispatcher. It is the same shift LLM serving already made — mixture-of-experts logic applied at the system level, where each expert is an entire specialized diffusion model.
07Efficiency - PixelRush

Big Images, Small Memory

<strong>Generate 4- and 16-megapixel images on a phone by refining the picture in patches of latent space, not all at once.</strong> Roughly 35x faster — minutes down to seconds.

It's maybe 35 times faster. You know, from let's say uh 10 minutes to uh around 20 seconds type of you know acceleration.

— Fatih Porikli, TWIML AI Podcast
Key Insight
The trick isn't a bigger model — it's refusing to scale the latent space with the output. Baselines grow the latent proportional to resolution and pay in memory and time; PixelRush keeps the latent at 1K size and tiles it. That makes the seams the hard problem, so the real innovation is where they inject noise: more at patch boundaries, less in the centers.
08Editing - InverField

Editing Without the Seams

<strong>Remove or replace an object by inverting the real photo into noise specific to that image — not random noise.</strong> The edit blends in with no telltale boundary artifacts.

So this noise is not random noise anymore. It is specific to the input image. So if I change the input image the noise is going to be different.

— Fatih Porikli, TWIML AI Podcast
Key Insight
Most inpainting starts the masked region from pure noise, which is why removed objects leave ghostly, mismatched patches. Inverting the whole image into its own noise means the edit is born already consistent with its surroundings — the background leaks into the fix by construction, instead of being stitched on afterward.